1 Introduction

For more than a decade, the people of northern Thailand have suffered from air pollution caused by open agricultural burning and forest fires during dry and hot weather conditions from January to May, frequently called the fire season or burning season. During this time, local villagers set fires to burn rice straw or weeds in order to clear land, remove weeds, and control disease before starting new crops. Forests and underbrush are also burned to enhance the reproduction and propagation of natural crops such as mushrooms or the edible plants Sauropus androgynus and Melientha suavis. These fires produce large amounts of small particulate matter (PM) in the northern provinces of Thailand (Chiang Rai, Chiang Mai, Mae Hong Son, Lampang, Lumphun, Phayao, Nan, Phrae, and Uttaradit). These aerosols are a serious threat to public health and tourism (Jeensorn et al. 2018; Silva et al. 2016; Air Quality and Noise Management Bureau of Thailand 2013; Ryan et al. 2013). However, the local staff and budget available to deal with air pollution management are limited in Thailand and most developing countries; thus, identifying and prioritizing critical areas (those experiencing repeated burning and unhealthy concentrations of PM10, i.e., PM with a diameter < 10 μm) would be beneficial.

For many decades, satellite-based remote sensing has been a useful tool for mapping changes in land use/land cover (LULC). Furthermore, many remote sensing applications have been developed for retrieving aerosol optical depth (AOD) or aerosol optical thickness (AOT) air quality measurements using moderate- to low-resolution data from the MODerate-resolution Imaging SpectroRadiometer (MODIS) or Multi-angle Imaging SpectroRadiometer (MISR) aboard the Terra satellite or various Landsat products. The derived AOD/AOT or atmospheric reflectance can then be compared to PM2.5 (PM with a diameter < 2.5 μm) or PM10 data acquired from other instruments such as the Aerosol Robotic Network (AERONET), physical measurements (such as sun photometers, spectroradiometers, and PM laser photometers), or local air quality measurement stations (Dumitrache et al. 2016; Meng et al. 2016; Hagolle et al. 2015; Benas et al. 2013; Bhaskaran et al. 2011; Othman et al. 2010; Hadjimitsis 2009). Remote sensing PM10 data enable the estimation of air pollution levels over large and remote areas. AOD images from MODIS are widely used to develop PM2.5 and PM10 estimation algorithms at global or regional scales, while Landsat images are more suitable for local or provincial studies (Roy et al. 2017; Saraswat et al. 2017; Shaheen et al. 2017; Saleh and Hasan 2014; Mishra et al. 2012). However, the AOD/AOT retrieval algorithm for PM estimation from Landsat is not as well developed as it is for MODIS (Sun et al. 2016; Glantz and Tesche 2012; Remer et al. 2005; von Hoyningen-Huene et al. 2003).

In this study, we used multi-temporal Landsat 8 imagery acquired during the 2015–2017 fire seasons in Nan Province, northern Thailand, to analyze PM10 concentrations and land use. The estimated PM10 concentrations were then classified to health risk levels based on criteria defined by Thailand’s Air Quality and Noise Management Bureau, Pollution Control Department, Ministry of Natural Resources and Environment. Areas with high PM10 levels were also correlated with LULC type (Vadrevu et al. 2018; Zahari et al. 2016; Zou et al. 2016b; Wu et al. 2012), and the locations of biomass burning or fires (hotspot data from the MODIS sensor) acquired during the 2015–2019 fire seasons. This enabled identifying and prioritizing critical areas experiencing repeated burning and unhealthy PM10 levels for improved air pollution management. These findings will be useful for many local agencies tasked with air pollution control, fire management, forest conservation, disaster prevention and mitigation, public health, and land use or agricultural planning. Such agencies can use the resulting data as a resource for deciding where to focus local efforts and for developing strategic plans for air pollution management.

2 Materials and methods

2.1 Study area

Nan Province is bordered by many mountainous forested national parks, with a narrow swath of flat terrain in the middle of the province. Almost 900 villages are located in 99 sub-districts (Tambon) within 15 districts (Amphoe). Most villagers work in the agricultural sector growing maize, rice, rubber, tree fruits (longan, mango, lychee, tamarind, lambutan, etc.), teak, and other crops. Agricultural land use in Nan has increased for at least the past decade, making biomass burning an increasingly critical issue in this area (Nan Statistical Office Annual Report 2015–17). Figure 1 presents a map of the study area showing the provincial boundary of Nan (highlighted boundary) and neighboring provinces where eight of the PM10 monitoring stations used in this study were located (all in the same Landsat scene; path–row 130–47) as well as the main roads in Nan. In addition, the industrial land consisted of agricultural product (dry longan, tobacco, rice mill) and teak processing plant in Nan was overlaid. Small-scale or household industrial land uses cover less than 2% of the entire areas in Nan; hence, these small portions cannot be seen clearly on the map. These components were overlaid on the mosaicked scenes from Landsat 8 (Path–Row: 129–47, 130–46, and 130–47) acquired in 2015.

Fig. 1
figure 1

Study area

Full size image

2.2 Data

The details of all imagery, air quality data, and other data used in this study are given in Table 1. We used three Landsat 8 scenes to cover all of Nan, as 70% of the area was within Path–Row 130–47, as well as most ground stations (see Fig. 1). In addition, cloud-free images are rare in Southeast Asia or the tropical climate zone; thus, some cloudy or smoky scenes were initially included. After defining the PM10 concentration maps, images that had high error levels were omitted to avoid overestimating unhealthy areas (Table 2). Unfortunately, most of the Landsat 8 operational land imager (OLI) images acquired during January–May 2018 had serious cloud contamination problems; therefore, we could not produce PM10 concentration maps for 2018.

Table 1 Data sources
Full size table
Table 2 Specific Landsat 8 scenes used
Full size table

Thailand has a limited number of AERONET member stations, the closest of which is located in Chiang Mai Province, almost 300 km from Nan (NASA 2018). Therefore, in situ PM10 time series data were collected from eight Thai Pollution Control Department air quality measurement stations located in Nan and neighboring provinces (see Fig. 1). A few meteorological stations are located in Muang Nan (almost 10 km from the PM10 monitoring station located in the central of Muang Nan), Tha Wang Pha, and Thung Chang. However, the observations in these stations have limited parameters and are aimed at agricultural activities. Because of the uncomplete historical data and less significant weather fluctuations during our observed period, the burning season (dry weather), these meteorological parameters were not included in our analysis.

2.3 Methods

The study methodology is presented in Fig. 2. First, the selected Landsat data were processed for radiometric calibration by converting the digital number (DN) of the multi-spectral OLI data (using no thermal bands; Table 3) to the top-of-atmosphere (TOA) radiance using the radiance rescaling factors in the Landsat metadata (MTL) file as follows (Vermote et al. 2016):

$$L_{\lambda } = M_{L} \times Q_{\text{cal}} + A_{L}$$
(1)

where \(L_{\lambda }\) is the TOA spectral radiance (W m−2 sr−1 μm−1), ML is the band-specific multiplicative rescaling factor from the metadata (Radiance_Mult_Band_x, where x is the band number), AL is the band-specific additive rescaling factor from the metadata (Radiance_Add_Band_x), and Qcal is the quantized and calibrated standard product pixel values (DN).

Fig. 2
figure 2

Methodology flowchart

Full size image
Table 3 Landsat 8 OLI bands used in this study
Full size table

The multi-spectral radiance images were then processed for atmospheric correction using the Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) algorithm in the Environment for Visualizing Images (ENVI) application (ExelisVisual Information Solutions, Inc. 2013). FLAASH incorporates the MODerate-resolution atmospheric TRANsmission (MODTRAN) radiation transfer code and works well with wavelengths from those in the visible light range up to 3 µm. Each image was computed based on the specific standard MODTRAN model atmospheres and aerosol types, with surface reflectance (SR) bands as the output:

$$L = \frac{A\rho }{{1 - \rho_{e} S}} + \frac{{B\rho_{e} }}{{1 - \rho_{e} S}} + L_{a}$$
(2)

where ρ is the pixel SR, ρe is an average SR for the pixel and surrounding region, S is the spherical albedo of the atmosphere, La is the radiance backscattered by the atmosphere, and A and B are coefficients that depend on atmospheric and geometric conditions, respectively, but not at the surface.

The difference between the derived TOA and SR is the path radiance (PR), which represents noise or radiance that is backscattered by particles and molecules in the atmosphere. Therefore, it was used to estimate the PM10 concentration in the air (Fernández-Pacheco et al. 2018; Roy et al. 2017; Yoram 1993). The PR images were correlated with PM10 (µg m−3) acquired on the same date and time at the air quality measurement stations. The correlation between each PR image (independent parameter) and PM10 from the stations (dependent parameter) when compared at the same geographic location was calculated as follows:

$$Y = aX + b$$
(3)

where a is the gradient slope, b is the y intercept, and X is the variable. The R2 of each band in each PR image varied from less than 0.1 to 0.86. We selected only the linear regression equations from bands/dates with high correlations (R2 ~ 0.7 or higher) to further estimate PM10, such as the correlation between Band 1 (X1 or PR derived from the Band 1/ultra blue (coastal/aerosol)) acquired on February 28, 2015, and the PM10 values measured at ground stations on the same date. The linear regression with the highest R2 (0.85,885) was used for PM10 prediction; the following derived linear Eq. (4) was applied to the image for PM10 estimation:

$$Y = 0.00004X + 0.01224$$
(4)

However, higher correlation coefficients were found for different bands on different dates. Thus, the relationship between several independent parameters or bands of PR images (multi-spectral and multi-temporal) and PM10 values measured from ground stations in the study area (dependent parameters) was analyzed using a multiple linear regression equation:

$$Y^{\prime} = a_{0} + a_{1} X1 + a_{2} X2 + \cdots + a_{n} Xn$$
(5)

where Y′ is the predicted PM10 (dependent variable), X1 through Xn are predictor variables (Table 3), a0 is the value of Y′ when all independent variables are equal to zero, and a1 through an are the estimated regression coefficients. The best subset for use in the forecasting model for estimating PM10 was analyzed using the stepwise method that either began with no variables in the model and added one variable at a time (called forward regression) or began with all potential variables in the model and removed one variable at a time (called backward regression). In this study, the best subset of variables used in the forecasting model for the estimated PM10 comprised X1, X4, and X7, which had R2 = 0.704 (RMSE ~ 17 µg m−3).

The derived linear regression equations were applied to PR image(s) to predict PM10 values, which were compared to the PM10 measured at ground stations. The differences between the predicted and measured PM10 values were converted to percentages. Our predicted PM10 values showed less than 20% error and were then reclassified into five health impact levels (Table 4) based on guidelines from Thailand’s Air Quality and Noise Management Bureau (2013), which in turn refer to US Environmental Protection Agency (1999) guidelines.

Table 4 Air quality classifications by PM10 level, based on Thai government standards
Full size table

PM10 data were consolidated by level and overlaid with LULC data derived from Landsat 8 imagery from January 27, 2015. The study area contained five main LULC classes: agricultural land (paddy fields, orchards, and crops), which covered 61.90% of Nan Province, dense forests (thickly treed and difficult to access by road) covered 14.75%, disturbed forests (treed but mostly disturbed or accessed by human activities) covered 19.03%, developed land and other (construction and areas not matching the other classes) covered 3.99%, and water bodies covered 0.33% of the entire province. For each main class, we homogeneously defined sub-classes for supervised maximum-likelihood classification. The classification results were validated using ground-truth data from 2015, while the forest status was rechecked using field data from 2016 and 2017. Table 5 presents the accuracy and error percentages (error of omission and commission) of each LULC class in the confusion matrix. Every class exceeded the expected 70% accuracy, although disturbed forests (accuracy = 78.85%) were confused with agricultural land (error of omission = 10.38%, error of commission = 13.46%) and dense forests (error of omission = 12.90% and error of commission = 7.69%). The accuracy of the classified images was high (overall accuracy = 85.42%, Kappa coefficient = 79.85%); thus, they were used for a correlation analysis in the next step.

Table 5 Accuracy assessment of the LULC 2015 in Nan Province classified from Landsat 8 OLI
Full size table

The active fire hotspot data represent geographic coordinates with strong emissions in the mid-infrared (4 μm and 11 μm) wavelengths, as detected by the MODIS sensor on the Terra satellite in the morning and on the Aqua satellite in the afternoon. However, some pixels had higher temperatures than surrounding pixels, although no fire was present (these so-called “artificial fires” included metal roofs or other areas that appear to be very bright). The confidence of fire occurrence should therefore be considered when using this product (Junpen et al. 2013). We selected hotspot data during 2015–2019 with > 80% confidence values, and we correlated these real fire pixels with the LULC and PM10 concentration maps. The occurrence of more frequent fires and unhealthy PM10 concentrations implied more critical environmental problems (fires and air pollution) in an area and represented problematic human practices such as slash-and-burn techniques (converting forests to agricultural lands, burning fields before starting new crops) and adverse health impacts on local residents. Finally, critical areas were identified for management prioritization at the district level.

3 Results

3.1 PM10 concentrations

Only six PM10 images with ≤ 20% error were classified by air quality (Fig. 3); the area of each classification is summarized in Table 6. In January 2015, good air quality (9566 km2; ~ 78%) occurred over most of the study area, especially in mountainous areas, while moderate air quality (2638 km2; ~ 22%) was found in the lower central region of the province. In February 2015, unhealthy (2854 km2; ~ 23%) and very unhealthy (902 km2; ~ 7%) air qualities were centered in southern mountainous areas including the Wiang Sa, Na Noi, and Na Muen districts, caused by southerly dry season flow and increased burning. In March 2015, unhealthy air quality (11,230 km2; ~ 92%) dominated the province along with areas of very unhealthy air quality (974 km2; ~ 8%). In February 2016, moderate air quality (9798 km2; ~ 80%) was dominant, with unhealthy air quality (2406 km2; ~ 20%) occurring in some central and mountainous western regions including the Wiang Sa, Na Noi, Na Muen, and Mueang Nan districts. In February 2017, good (8799 km2; ~ 72%) and moderate (3406 km2; 28%) air qualities dominated. In March 2017, unhealthy (7223 km2; ~ 59%) and very unhealthy (2379 km2; ~ 19%) air qualities were most common, including in the Wiang Sa, Na Noi, Mae Charim, Mueang Nan, and Na Muen districts. Unhealthy and very unhealthy air qualities were predominant in Wiang Sa District for all dates, followed by the Na Noi district.

Fig. 3
figure 3

Monthly air quality maps during the 2015–2017 burning seasons in Nan Province, Thailand

Full size image
Table 6 Spatial coverage (%) of air quality levels on certain days (see Fig. 3) during the burning seasons in Nan Province, Thailand
Full size table

3.2 Relationship between PM10 concentrations, LULC, and hotspots

We selected areas classified as unhealthy and very unhealthy during February 2015, March 2015, February 2016, and March 2017 and calculated their overlays with LULC types (Table 7). It was found that > 60% of unhealthy and very unhealthy areas in the observed periods were related to agricultural land, < 20% to forest, and < 4% to other classes.

Table 7 Correlation between unhealthy and very unhealthy air quality levels and LULC types by observation
Full size table

During the 2015–2019 burning seasons, 1515 hotspots were identified in the study area (695, 294, 147, 118, and 297 in 2015, 2016, 2017, 2018, and 2019, respectively). Only ~ 1–2% of the total hotspots were observed in January or February, but this percentage increased drastically in March (in 2015–2019, ~ 42%, ~ 22%, ~ 40%, ~ 30%, and ~ 31%, respectively). The largest number of hotspots was observed in April (~ 56%, ~ 69%, ~ 48%, ~ 52%, and ~ 26% in 2015, 2016, 2017, 2018, and 2019, respectively), but their occurrence fell significantly in May (~ 0%, ~ 5%, ~ 9%, ~ 2%, and ~ 10% in 2015, 2016, 2017, 2018, and 2019, respectively). Overall, ~ 67% of the total hotspots occurred over agricultural land, ~ 16% over disturbed forest, ~ 13% over dense forest, and ~ 2% over developed land (Table 8). In 2015, the largest number of hotspots occurred on agricultural lands (~ 91%) and forest areas (~ 5% in disturbed forest and ~ 2% in dense forest) in the Wiang Sa, Chaloem Phra Kiat, Bo Kluea, Pua, Na Noi, Na Muen, Thung Chang, and Mae Charim districts, with similar patterns in 2016 and 2017, although with lower overall totals (Fig. 4). In 2018 and 2019, the largest number of hotspots occurred followed a similar pattern as that in 2015 to 2017 on agricultural lands (~ 64%) and forest areas (~ 16% in disturbed forest and ~ 16% in dense forest). The number of hotspots in 2019 was close to the number that occurred in 2016, which was higher than the number in 2017 and 2018, while more hotspots were distributed in agricultural land and forest areas than in the previous years. Overall, Wiang Sa had the most fires, followed by the Pua, Na Noi, and Chaloem Phra Kiat districts.

Table 8 Number of hotspots observed during the burning season by LULC category
Full size table
Fig. 4
figure 4

2015 LULC classifications and MODIS-derived fire hotspots during 2015–2019 in Nan Province, Thailand

Full size image

Unhealthy and very unhealthy PM10 concentrations began to cover the southern and central parts of the study area in February, especially in the Wiang Sa, Na Noi, Mae Charim, Mueang Nan, and Na Muen districts. This expanded to the entire study area in March due to southerly winds and new burning sites (Thai Meteorological Department 2015). Burning hotspots reoccurred most commonly on agricultural lands in the central and northern parts of Nan Province. Although hotspots in disturbed and dense forests were much less common, they ranked second and third in terms of reoccurrence.

Based on information from Thailand’s Air Quality and Noise Management Bureau, biomass burning and unhealthy PM10 both peak in March every year. During this time, Wiang Sa and Na Noi were the most critical districts for air pollution according to the results our study. Several villages in the Wiang Sa District, namely Naka Village in the Yai Hua Na Sub-district and Wana Prai and Nam Pi villages in the Nam Muab Sub-district, were located near highly repetitive burn sites and should be prioritized for air pollution management by collaborating with local villagers. Specifically, villagers should be encouraged to adjust their practices and behavior by stopping open burning, creating fire barriers (open buffer zones between forest and other land uses), preserving forests, monitoring and reporting fires to the appropriate agency, and taking greater personal care for their health during unhealthy air pollution conditions, among other measures.

4 Discussion

In this study, the correlation between three PR images calculated from the coastal aerosol, red, and SWIR2 bands of Landsat 8 and ground station-based PM10 levels presented the highest correlation and lowest error. By contrast, Saleh and Hasan (2014) showed that the coastal aerosol, blue, and green bands of Landsat 8 showed the highest correlation and lowest error. The difference between our predicted and measured PM10 values was < 20%, which was acceptable for determining PM10 concentrations in different LULC types. However, our algorithm should be further developed in order to reduce errors that were mainly caused by the limited availability of cloud-free images and ground measurement data. Kanniah et al. (2016) suggested integrating aerosol products from various satellite sensors to overcome the cloudiness (a common phenomenon in Southeast Asia) and orbital gaps of satellite tracks. In many studies, meteorological parameters (atmospheric pressure, relative humidity, air temperature, wind direction and speed, etc.) have provided valuable support for enhancing the prediction of seasonal PM concentrations and dispersion patterns at the regional scale when they are incorporated with daily AQI data (which have high temporal resolution/high frequency) and AOD (high temporal resolution/high frequency but low spatial resolution) from satellites. Nonetheless, the prediction results have shown very random affects (Diao et al. 2019; Qiao et al. 2019; Kliengchuay et al. 2018; Kamarul Zaman et al. 2017; Chu et al. 2016; Yin et al. 2016). These parameters less effectively enhance PM10 prediction when low-temporal-resolution data or long gap satellite observations (medium to high spatial resolution) are incorporated at the local scale when the relative magnitude of meteorological parameters seem to be similar. In our case, not only were limited parameters recorded from the meteorological station in Nan Province, but also, we focused on PM10 concentration only in the burning season (January to May), which is a period of dry weather. Therefore, these meteorological parameters were not incorporated into our analysis. Land use information has also been mentioned in some reports as a necessary parameter for improving the reliability of predicting PM10 concentrations over different LULC or for different sources of pollution (Meng et al. 2016; Zahari et al. 2016; Zou et al. 2016b). In addition, in most of the previous studies, PM10 comes from incomplete combustion, automobile emissions, or dust from big cities, industrial areas, coalmines, and arid areas. By contrast, the major source of PM10 in this study was evidently biomass burning in forest areas and agricultural lands, and the locations of these fires and sources of PM10 were random or could not be fixed). Although LULC was not combined into our prediction model, it was incorporated into the PM10 concentration maps and hotspots in our analysis for identifying priority air pollution management areas.

More advance and complex models (nonlinear regression methods) have been developed for predicting PM (both PM2.5 and PM10) using relevant covariates including high-temporal-resolution (e.g., daily) AOD products, meteorological parameters, land use information, or any related auxiliary data. For example, generalized additive model (GAM), which is an extension of multiple linear regression models, linear mixed-effects (LME) or mixed-effect model (MEM), geographically weighted regression (GWR), geographically and temporally weighted regression (GTWR), artificial neural network (ANN), support vector regression (SVR), and random forest (RF) have all been applied. Many studies compared the ability of the resulting developed models with the simple and widely used linear regression (Park et al. 2019; Li et al. 2018; Abdullah et al. 2019; Guo et al. 2017; Jiang et al. 2017; Zou et al. 2017; Kamarul Zaman et al. 2017; Chu et al. 2016; Nguyen et al. 2014). These additive or newly developed nonlinear models have presented higher correlation than linear regression when dealing with long-term and high-frequency data such as AOD/PM from satellites and ground-based PM observation data, daily meteorological parameters, LULC or source of pollution, and other related data. However, it has been difficult to conclude which prediction model is the best as their results have shown medium to high accuracies for large areas or at the national scale. However, many works have found that the linear regression method could still be used to predict PM concentrations, especially when only two parameters were considered, PM data from satellites (i.e., AOD, PR) and ground measurements. PM10 concentrations have been found to increase linearly with the increase in PR or AOD, and the prediction results from the simple linear method presented medium to high accuracy (Roy et al. 2017; Shaheen et al. 2017; Nguyen et al. 2015; Glantz and Tesche 2012; Mishra et al. 2012; Themistocleous et al. 2012; Othman, et al. 2010). Our multi-spectral model accurately (high R2) predicted PM10 concentrations in the study area (R2 = 0.704, RMSE ~ 17 μg m−3). In addition, our model shows a similar accuracy to the one derived using nonlinear methods in previous studies that used the same number of or more input parameters (R2 = 0.3–0.99, RMSE = 8–31 μg m−3).

Classifying the estimated PM10 concentrations into the five classes defined by the Thai government (and based on US EPA guidelines), we showed that agricultural land, disturbed forest, and dense forest were repeatedly subjected to unhealthy levels of PM10 concentrations (> 120 µg/m−3) during the burning seasons of 2015–2017. Using different guidelines, Roy et al. (2017) found that approximately half of Vadodara city and Nandesari town in India showed unhealthy levels of PM10 concentrations (120–160 µg/m−3, with some small areas at 160–200 µg/m−3). However, they did not consider specific LULC types or extents. In the Gaza Strip, Shaheen et al. (2017) used two major land classes (urban and non-urban) to assess pollution levels at 28–144 µg/m−3. Most of their study area had good to moderate air quality. However, it is difficult to compare our results to those of studies that are presented for AOT or PM10 concentrations in graduated colors and values (not in µg/m−3), such as those reported by Themistocleous et al. (2012).

Although many studies have estimated PM10 concentrations using satellite data (such as Landsat 8), to the best of our knowledge, this is the first time fire hotspots and LULC types have been integrated into PM10 estimation to prioritize air pollution management efforts.

5 Conclusions

Our empirical model based on ground station data estimated PM10 concentrations with satisfactory accuracy during the burning season (January to May) in Nan Province, Thailand. The 20% error rate was acceptable for the purpose of broadening our perspective on PM10 concentrations and their relationship with LULC and fire hotspots in the study area. Most problematic PM10 concentrations and fire hotspots were associated with agricultural land, disturbed forest, and dense forest areas. Locations with repeated occurrences from 2015 to 2019 showed especially strong correlations with agricultural land and forests.

As local agencies in Thailand and other developing countries face budgetary and staffing limitations, defining the most-affected areas using this method enables focusing attention and resources to the greatest possible effect. For example, agricultural and natural resources offices can use such results (locations of high air pollution density or high-frequency hotspots) for controlling or mitigating open-burning activities on agricultural land and monitoring potential forest fires. Moreover, pollution control, disaster prevention and mitigation, and public health offices can better cooperate with local villagers and farmers to properly adapt and respond to these issues.

The algorithm for PM10 or AOT estimation presented here could be improved by (1) increasing/updating the number of high-quality samples, more cloud-free images with medium to high spatial resolution from various observation sensors (Park et al. 2019), PM10 measurements from ground stations or reliable handheld instruments to obtain a more reliable correlation analysis (e.g., well-distributed sites with frequent records in the study area), and incorporating other covariates such as weather and climate (increased local observation source/data) that influence the amounts and patterns of PM10 concentration and distribution in a study; (2) testing various types of regression algorithms or ensemble models that are applicable to the local or provincial scale (Shtein et al. 2020; Zhang et al. 2019; Zou et al. 2016a); (3) investigating other methods for aerosol retrieval or atmospheric correction using Landsat data or data with higher spatial resolution (Diao et al. 2019); and (4) considering other air pollutants that are derived from biomass burning such as SO2 and NO2 (Qiao et al. 2019).

The low temporal resolution of Landsat 8 results in satisfactory mapping of LULC types and PM10 concentrations at local to regional levels, but it limits our ability to observe and predict the pattern and direction of PM10 dispersion in near real time. Although higher-temporal-resolution products (such as those from MODIS) have better capabilities for near-real-time air quality prediction at national to global levels, they also have poorer spatial resolutions that are not suitable for identifying LULC types.