Estimation of fine particulate matter in an arid area from visibility based on machine learning

Abstract

Background

The absence of air pollution monitoring networks makes it difficult to assess historical fine particulate matter (PM_2.5) exposures for countries in the areas, such as Kuwait, which are severe impacted by desert dust and anthropogenic pollution.

Objective

We constructed an ensemble machine learning model to predict daily PM_2.5 concentrations for regions lack of PM_2.5 observations.

Methods

The model was constructed based on daily PM_2.5, visibility, and other meteorological data collected at two sites in Kuwait. Then, our model was applied to predict the daily level of PM_2.5 concentrations for eight airports located in Kuwait and Iraq from 2013–2020.

Results

As compared to traditional statistic models, the proposed machine learning methods improved the accuracy in using visibility to predict daily PM_2.5 concentrations with a cross-validation R² of 0.68. The predicted level of daily PM_2.5 concentrations were consistent with previous measurements. The predicted average yearly PM_2.5 concentration for the eight stations is 50.65 μg/m³. For all stations, the monthly average PM_2.5 concentrations reached their maximum in July and their minimum in November.

Significance

These findings make it possible to retrospectively estimate daily PM_2.5 exposures using the large-scale databases of historical visibility in regions with few particulate matter monitoring stations.

1. Introduction

Populations in arid areas, such as Kuwait, are exposed to high air pollution levels due to large anthropogenic emissions and dry climate ^{1, 2}. Exposures to fine particulate matters (PM_2.5) have been closely associated with health effects worldwide, such as cardiovascular disease, respiratory disease, brain health, lung cancer, and premature mortality ^3–8. Unfortunately, there is a general scarcity of integrated air pollution monitoring networks in arid areas, which makes it difficult to assess historical PM_2.5 exposures.

To date, a few PM_2.5 measurement studies have been conducted in Kuwait and other Middle Eastern countries ^9–14. These studies have reported high levels of particulate air pollution contributed, which they attributed to various local and regional anthropogenic factors. The annual PM_2.5 concentrations were found to exceed the World Health Organization guidelines and the U.S. National Ambient Air Quality Standards. However, there remains a lack of continuous historical PM_2.5 measurement data.

As a routine, visibility data is collected either at airports or meteorology stations around the world. The relationship between visibility and PM_2.5 is relatively stable due to the light extinction (scattering and absorption) effects of particles with similar sizes to the wavelengths of visible light, particularly in those arid areas ^15–18. Some studies relying on visibility to estimate historical ground-level PM_2.5 exposures in the absence of PM_2.5 data collected through observation ^18–20. Based on this approach, large-scale databases of historical visibility data collected by airport or meteorology stations can be used to address the lack of PM_2.5 exposures.

The relationship between visibility and PM_2.5 concentrations may be non-linear and can be affected by many physical factors, such as relative humidity²¹. However, most previous studies have used linear statistic models, such as generalized linear model or linear mixed effects model, to quantify the visibility-PM_2.5 relationship ^{18, 19, 22}. Machine learning methods can be used to model nonlinear and complex relationships, however, they have not been used to quantify visibility-PM_2.5 relationships to date.

In this study, we collected daily PM_2.5 samples and visibility data for two residential locations in Kuwait. We developed an approach to ensemble machine learning to predict PM_2.5 using visibility and other parameters, using the large-scale database of historical visibility in a region without PM_2.5 monitoring networks.

2. Methods

2.1. PM_2.5 Sampling and Analyses

The data used for the machine learning model proposed in this study was collected in Kuwait, a small desert country located in the north of the Persian Gulf, to the south of Iraq, and northeast of Saudi Arabia, with typical climatic conditions representative of the Middle East. It is frequent for severe dust storms to strike Kuwait, especially during the spring and summer months ².

We deployed PM_2.5 samplings at two locations in Kuwait, as shown in Figure 1. The first sampling site, a rooftop in the Kuwait University (29.3217° N, 47.9719° E), is an urban residential area at the heart of Kuwait City (“Site1”). It is less than 3 km away from the city center. The second site is a rooftop at the local healthcare center in the Ali Sabah Al-Salem area (28.9570° N, 48.1548° E). This is a residential area located in the south, about 50 km away from Kuwait City (“Site2”). This area is surrounded by oil fields, industrial premises, and a sewage treatment plant. The elevations at Site 1 and Site 2 were nearly 18 and 26 meters above sea level, respectively. We collected daily PM_2.5 concentrations at both sites during the period from October 2018 to December 2019.

Figure 1.

Map of the two sampling sites in Kuwait (Black stars).

We employed the custom-designed Harvard Impactor (HI) samplers capable to collect the particles in large amounts, especially in case of a dust storm. Conventional sampling devices are equipped with impaction substrates that are prone to particle bounce, which reduces sampler accuracy ¹¹. The HI sampler uses a slit acceleration jet and polyurethane foam as an impaction substrate. This substrate has a high capacity for removing and holding particles larger than cut-points during dust storms. Polyurethane foam has a capacity of 35 mg of particles, which allows us to collect particles in challenging desert environments ²³. The HI sampler was previously field tested in Kuwait City by Brown, Bouhamra, Lamoureux, Evans and Koutrakis ⁹. PM_2.5 filter samples from two collocated samplers were collected on a pre-weighed Teflon filter for 24 hours between 9:00 am and 9:00 am the next day. Gravimetric analyses were conducted at the Harvard School of Public Health using a Mettler MT-5 microbalance in a particle-free environment at controlled temperature (18–24°C) and relative humidity (35–45%). The average of the two daily samples was used for analysis. Due to logistic restrictions, there were consecutive missing days in June and July of 2018 and sampling once every other day in January and July of 2019. We removed samples when there was a power failure in the devices (n=2), or when the PM_2.5 concentration was higher than the PM₁₀ concentration on the same day at the same location (n=3) during the daytime without any recorded dust events or local emissions.

We also measured hourly visibility and other meteorological parameters, including relative humidity, average wind speed, temperature, dew point temperature, and atmospheric pressure at the same location as the PM_2.5 data. The hourly data were averaged over a 24-hour period to match daily PM_2.5 measurements.

2.2. Machine Learning Approaches

We assessed the predictive ability of three machine learning models in our study, including the neural network model, random forest model, and gradient boosting model. All of these three machine learning models were applied to model the complex relationship between the dependent variable and predictor variables with the assistance of different algorithms. The neural network is based on a multi-layer feedforward artificial neural network (ANN) that is trained with stochastic gradient descent using back-propagation. The random forest is a powerful machine learning method for classification and regression, which generates a forest of classification or regression trees, rather than just a single tree. Gradient boosting is a forward learning ensemble method that provides good predictive results based on increasingly refined approximations. The three algorithms have all been successfully used to estimate PM_2.5 concentrations ^{20, 24, 25}. The details of these machine learning models can be found in the work of Bishop ²⁶. The variables used in each model include visibility, relative humidity, wind speed, dew point temperature, atmospheric pressure, month, and day of the week. A list of the key hyperparameters for each machine learning model was provided in Table S1. Then, an ensemble machine learning method was applied to integrate the three models for improving the performance. The ensemble method is a supervised ensemble machine learning algorithm with the capacity to identify the optimal combination of a range of prediction algorithms through a process known as stacking ²⁷. The equation below describes the ensemble model:

$${\text{PM}}_{2.5}={\text{f}}_1\left({\text{PM}}_{2.5}\right)+{\text{f}}_2\left({\text{PM}}_{2.5}\right)+{\text{f}}_3\left({\text{PM}}_{2.5}\right)$$

Where f₁, f₂, f₃, and f₄ represent the predictions made using the neural network, the random forest, and the gradient boosting, respectively.

A 10-fold cross-validation was performed by randomly dividing all of the visibility data into 10–90% splits (10 random groups). For each split, the model was trained with data using 90% of the visibility data to predict PM_2.5 concentrations for the 10% of visibility data excluded from the training. This process was repeated 10 times, and the cross-validation R² (CV R²) was calculated.

Since 10-fold CV was not sufficient to validate the model performance for long-term predictions at a different site, we conducted more validation approaches to address this problem. In the first step, we fitted the ensemble model based on one site and used it to predict concentrations at the other. In the second step, we fitted the ensemble model by leaving out the observation in October, 2018, and validated the model by the observations in October, 2018. We did the same validation for the data in October, 2019. The model was carried out in R software using the H₂O package (http://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/stacked-ensembles.html?highlight=ensemble).

3. Results

3.1. PM_2.5 and visibility levels

There were 554 matched pairs of daily PM_2.5 and visibility reading obtained. Figure 2 shows the daily PM_2.5 concentrations from the two measurement sites during the study period. Table 1 lists the averages of PM_2.5 concentrations and meteorological parameters at the two sites. The PM_2.5 levels (mean ± standard) at sites 1 and 2 were 33.98±15.86 and 44.66±18.14 μg/m³, respectively. There was a considerable variation in daily PM_2.5 concentrations, which were consistently higher at Site 2 than at Site 1. Throughout the study period, PM_2.5 levels exceeded the daily WHO standard (15 μg/m³) for 74% of the time at Site 1 and for 90% of the time at Site 2²⁸_. The correlation coefficient between the daily PM_2.5 concentration and the visibility was −0.56 for the two sites during the study period, with p values falling below 0.001, as shown in Figure 3.

Figure 2.

Daily PM_2.5 concentrations (μg/m³) for the two sampling sites in Kuwait during the study period.

Table 1.

PM_2.5 concentrations and the levels of meteorological data (mean ± standard deviation) at the two sampling sites from October 2018 to December 2019.

Variables	Site1 (N=239)	Site2 (N=315)
PM2.5 (μg/m3)	33.98±15.86	44.66±18.14
Visibility (m)	18,144.06±34,44.66	16663.76±4113.78
Relative humidity (%)	37.50±20.92	40.49±21.94
Weed speed (m/s)	3.09±1.09	2.72±1.09
Maximum weed speed (m/s)	5.06±1.81	4.43±1.61
Dew point temperature (°C)	9.61±4.73	11.03±6.16
Atmospheric pressure (kPa)	100.40±0.82	100.60±0.77

Footnote: N is the number of observing days for each site.

Figure 3.

Correlations between daily PM_2.5 concentrations (μg/m³) and visibility (meter) for the two sampling sites in Kuwait during the study period.

3.2. Model performance

Table 2 shows the 10-fold CV R² for the three machine learning models and the ensemble model. The CV R² obtained from the individual algorithms (deep learning, gradient boosting, and random forest) is 0.59, 0.62, 0.64, respectively. The performance produced by the ensemble model was superior to that of any single algorithm with a CV R² of 0.68. In addition, a traditional linear model was applied to predict the daily PM_2.5 concentrations using the same variables as the machine learning models. The CV R² obtained from the linear model is 0.51, which is lower as compared to the machine learning methods. The R² from different validation approach for the ensemble model are listed in Table 3. The addition validation R² is comparable to the 10-fold CV R², which indicated that our model could be used to make long-term predictions at a completely different site or different time.

Table 2.

The 10-fold CV R² for different machine learning models.

Model	CV R2
Deep learning	0.59
Gradient boosting	0.62
Random forest	0.64
Ensemble	0.68

Table 3.

The R² from different validation approach for the ensemble model.

Training dataset	Validation dataset	R2
Site 1	Site 2	0.66
Site 2	Site 1	0.66
All observations except October,2018	October,2018	0.65
All observations except October,2019	October,2019	0.65

Figure 4 presents the relationship between the predicted and measured PM_2.5 concentrations with the ensemble model used. According to the results obtained from the ensemble model, the relationship between predicted and monitored PM_2.5 values shows a high level of consistency (R²=0.91). Figure 5 shows variable importance from the random forest model, which demonstrated that visibility is a reliable proxy for PM_2.5.

Figure 4.

Relationship between measured and predicted PM_2.5 from the ensemble model.

Figure 5.

Variable relative importance from the random forest model.

3.3. PM_2.5 concentration predictions

Our ensemble model was applied to predict PM_2.5 concentrations using the visibility and metrological data collected at 8 different airports across Kuwait by NOAA from 2013 to 2020. Figure 6 shows the spatial distribution of predicted PM_2.5 concentrations for each location by year. Table 4 listed the predicted PM_2.5 concentrations for each station by year. According to the results obtained by us, the average yearly PM_2.5 concentrations for the 8 stations is 50.65 μg/m³. KUWAIT INTL located near downtown of Kuwait has the highest yearly average PM_2.5 concentrations among the 8 stations, with a value of 61.54 μg/m³, whereas SALMY located in the desert has the lowest values (43.59 μg/m³). There were few changes in the yearly PM_2.5 concentrations from 2013–2020 for KUWAIT INTL. For the other stations, the PM_2.5 concentrations in 2019 and 2020 were lower than in previous years. Figure 7 presents predicted PM_2.5 concentrations averaged by month for the 8 locations. For all stations, the monthly average PM_2.5 concentrations reached their maximum in July and its minimum in November.

Figure 6.

Spatial distribution of yearly average predicted PM_2.5 concentrations (μg/m³) for 8 stations in Kuwait from 2013–2020.

Table 4.

Yearly average predicted PM_2.5 concentrations (ug/m³) for 8 stations in Kuwait from 2013–2020.

Station	2013	2014	2015	2016	2017	2018	2019	2020	Average
ABDALY	50.65	51.24	54.48	51.46	53.44	54.39	47.79	49.23	51.58
ABRAQUE MAZRAA	36.95	46.04	46.70	49.2	49.85	49.78	47.35	48.03	46.74
KUWAIT INTL	61.41	60.98	61.96	61.18	62.41	61.44	61.75	61.20	61.54
MANAGISH	54.78	54.15	62.98	64.4	61.52	58.36	55.23	55.52	58.37
MITRIBAH	53.64	40.13	35.89	47.8	46.61	46.82	41.22	42.4	44.31
SABRIYAH	44.09	46.88	43.95	53.22	51.19	51.28	46.83	47.16	48.07
SALMY	42.03	44.05	48.73	45.02	45.11	44.79	39.63	39.39	43.59
WAFRA	48.74	49.76	52.79	50.78	52.55	51.94	49.93	51.22	50.97
Average	49.04	49.15	50.93	52.88	52.84	52.35	48.72	49.27	50.65

Figure 7.

Monthly average predicted PM_2.5 concentrations (μg/m³) for 8 stations in Kuwait from 2013–2020.

4. Discussion

The absence of air pollution monitoring networks in Kuwait makes it difficult to assess the historical PM_2.5 exposures. Toward this end, we developed an ensemble machine learning approach to convert visibility data into PM_2.5 estimates based on the paired daily visibility and PM_2.5 measurement data collected. This made it possible to predict daily PM_2.5 concentrations using a large-scale database of historical visibility.

Compared with traditional regression models, the machine learning methods proposed by us improved the accuracy of using visibility to predict daily PM_2.5 concentrations with a CV R² of 0.68. The model was applied to predict PM_2.5 concentrations at 8 visibility stations in Kuwait. Our predicted PM_2.5 levels exhibited similar spatial and temporal distributions as the PM_2.5 measurements conducted in these regions ¹³. Finally, the predicted PM_2.5 concentrations were similar to those predicted previously using satellite-based aerosol optical depth (AOD) ²⁰.

Our study is subject to some limitations. First of all, we fitted the model based on the PM_2.5 data collected from two measurement sites in around one year in Kuwait. These limitations could introduce bias into our predictions. It is suggested that more PM_2.5 monitoring data should be collected from the regions with limited PM_2.5 data to improve the accuracy of PM_2.5 exposure assessment. Secondly, the relationship between PM_2.5 and visibility can be affected by more variables than those used in our model. Thirdly, compared with the traditional statistic model, the machine learning approach are good at handling multiple co-linear problems among input variables and has better prediction performance. However, over fitting may occur when using a machine learning approach.

To the best of our knowledge, this is the first machine learning model constructed to estimate PM_2.5 concentrations using visibility data. This study is contributory to assess the health effects of PM_2.5 exposures and formulating effective environmental policies for those regions with limited PM_2.5 measurement data.

5. Conclusions

In this study, an ensemble machine learning model was constructed to predict daily PM_2.5 concentrations according to the visibility data and daily PM_2.5 samples collected in Kuwait. The daily PM_2.5 concentrations in 8 stations in Kuwait from 2013–2020 were assessed by our model. The model proposed by us improved the accuracy of predicting daily PM_2.5 exposures through the large-scale database of historical visibility.

Impact statement.

The scarcity of air pollution ground monitoring networks makes it difficult to assess historical fine particulate matter exposures for countries in arid areas such as Kuwait. Visibility is closely related to atmospheric particulate matter concentrations and historical airport visibility records are commonly available in most countries. Our model make it possible to retrospectively estimate daily PM_2.5 exposures using the large-scale databases of historical visibility in arid regions with few particulate matter ground monitoring stations. The product of such models can be critical for environmental risk assessments and population health studies.

Data availability

All data included in this study are available upon request by contact with the corresponding author.