1. Introduction
Animal feeding operations (AFOs) are agricultural facilities where animals are kept and raised in a small area. While animal agriculture is a large driver of the economy, these farms also create a strain on the environment. The eastern region of North Carolina (NC) has become highly populated with thousands of AFO farms, primarily housing poultry and hogs [
1]. NC is ranked 4th in the United States for hog production and 1st in the United States for all poultry production due to the strong turkey production [
2]. During 2016, Duplin and Sampson Counties, which are in Eastern NC, contained over 4 million hogs and 100 million chickens [
3].
Ammonia (NH
3) emissions from AFOs account for more than half of the reactive nitrogen released into the environment [
4,
5]. The largest emission source of NH
3 into the atmosphere is from agricultural sources [
6,
7]. NH
3 plays a significant role in the biogeochemical nitrogen cycle and causes health and environmental effects [
8,
9,
10,
11,
12,
13,
14,
15,
16,
17]. The manure that is produced by the animals is frequently used as fertilizer and spread on crops throughout the farm, further releasing the pollutant into the atmosphere. NH
3 also has several human health effects including eye, nose and throat irritation, dizziness and headaches [
18,
19]. Once NH
3 is released into the atmosphere it can then deposit by either rainfall or dry deposition to regional waterbodies or it can react with other compounds in the area to create other pollutants and cause further harm [
14,
15,
16,
17,
20,
21,
22]. Gaseous NH
3 in the atmosphere contributes to the formation of airborne fine particulate matter (PM
2.5) through reactions with water vapor and other air pollutants, including oxidation products of sulfur dioxide (SO
2) or nitrogen oxides (NO and NO
2, or NO
X) [
23]. Ammonium compounds, including ammonium sulfates (NH
4HSO
4 and (NH
4)
2SO
4) and ammonium nitrate (NH
4NO
3), make up a large fraction of PM
2.5 [
24] (defined as particulate matter with an aerodynamic diameter of 2.5 microns or smaller). Elevated concentrations of PM
2.5 are associated with respiratory issues, heart problems and has been linked to premature death [
25,
26,
27]. Ammonia can be important in the nucleation of new particles [
28,
29]. These ammoniated particles scatter light, attenuating visibility, and can result in some atmospheric cooling [
30]. Particulate matter emissions are released through primary sources (e.g., power plants, transportation) [
31] as well as secondary sources (e.g., gas to particle conversion). However, in the United States, regulatory strategies to reduce PM
2.5 have focused on primary emissions of PM
2.5 and on reductions of SO
2 and NO
X emissions; however, the control of NH
3 emissions can also be effective for reducing concentrations of PM
2.5.
The North Carolina Clean Smokestacks Act has reduced sulfur dioxide emissions by 89% compared to 1998 emissions [
32] and the improved regulations on vehicle manufacturers has helped to reduce NO
x emissions. Because these reductions have already been made, any additional effort to reduce both SO
2 and NO
x emissions would become less cost effective. Thus, some studies have started to indicate that a reduction in ammonia emissions would be a more cost-effective solution in terms of long-term reductions in particulate matter concentrations [
33,
34,
35,
36,
37,
38,
39,
40]. However, ammonia is not a criteria pollutant. Moreover, any reduction in particulate matter must be measured at monitoring stations, leading to economic, political and logistical problems.
Recent studies in Europe using detailed chemical transport models and time resolved NH
3 emissions illustrate the strong nonlinearity between PM
2.5 and NH
3 emissions, and the reduction in NH
3 emissions significantly reduces PM
2.5 levels in both summer and winter periods [
33]. The impact of NH
3 emissions on PM
2.5 depends on meteorological parameters (e.g., temperature, relative humidity), the magnitude of the perturbation to NH
3 emissions and the abundance of particulate nitrate (NO
3−), gaseous nitric acid (HNO
3) and particulate sulfate (SO
4= and HSO
4−), which are the products of the oxidation of SO
2 and NO
x, two byproducts of combustion [
31,
33,
35]. Utilizing this foundational knowledge of key processes and an integrated approach of using satellite measurements (ammonia and aerosol optical depth) and ground-based measurements (meteorology and PM
2.5), an observational-based statistical model is developed to predict PM
2.5 concentrations in these rural regions in an effort to advance Earth system predictability. The abundance of NO
3T and SO
4T was not measured.
Measurements of PM
2.5 in rural regions (especially in intensively agricultural locations) of the US are in general not conducted and are therefore not available. Recent studies [
21,
41] have shown that increased ammonia from agriculture leads to increased PM
2.5. Therefore, the purpose of this work is to estimate PM
2.5 concentrations in ammonia-rich environments of Eastern NC using a combination of satellite and in situ data. Using satellite data will allow us to develop a method for predicting ground-level PM
2.5 concentrations in areas of high agricultural influence which normally do not have a ground-based measurement site. By utilizing satellite-derived ammonia concentrations and aerosol optical depth (AOD) along with meteorological parameters, we can predict, with reasonable certainty, PM
2.5 concentrations across NC. Moreover, a thorough review of the scientific literature did not provide any studies offering predictions of PM
2.5 in rural regions based on our process-based approach (i.e., coupling of satellite and ground-based measurements). However, a complex numerical air quality model such as the Weather Research and Forecasting-Community Multiscale Air Quality (WRF-CMAQ) or WRF-Chem Modeling System may also be utilized to predict PM
2.5 and NH
3 emission from agricultural land using these air quality models across various regions of the world [
42,
43,
44,
45].
3. Results
Multiple linear regression is a classical and well-known statistical technique to quantify the association of a variable (called the dependent variable) with several other variables (called independent variables). ANOVA or analysis of variance is a statistical technique to gauge the statistical significance and practical usefulness of the linear regression towards explaining the dependent variable. In this analysis, we implemented multiple linear regression and ANOVA to quantify the association of PM
2.5 (the dependent variable) with six independent variables: ammonia, AOD, T, P, WS and RH. Furthermore, we constructed the ANOVA table to study the statistical significance of this multiple linear regression model. Multiple linear regression models were created to assess the ability of remotely sensed data to predict ground-based PM
2.5 concentrations in rural areas that are usually characterized as having high agricultural activity and no ground-based PM
2.5 monitor. The models were created for Cumberland and Johnston Counties in NC (i.e., training data) and then validated from Duplin County NC (i.e., test data) against the ground-based measurements located in the county. Cumberland and Johnston Counties were chosen specifically because of their proximity to Sampson County, which has the highest agricultural activity in that state next to Duplin County, which was chosen as the validation site for this reason. Sampson County could not be used for validation because there is no ground-based PM
2.5 monitor in the county. After reviewing and testing different model inputs (e.g., a variety of meteorological parameters with data from each county of interest), a combination of Cumberland County and Johnston County data for the entire summer period were chosen:
where PM
2.5 is given in units of µg m
−3, Ammonia is the total atmospheric column loading of ammonia as retrieved from IASI observations in molecules cm
−2, AOD is the MODIS aerosol optical depth, T is temperature in °C, P is pressure in millibars, WS is wind speed in meters per second and RH is relative humidity in percent.
This model gave an r
2 value of 0.43. The specific statistics can be seen in
Table 2 (an analysis of variance or ANOVA table). Here, DF stands for “degrees of freedom”, i.e., the number of values that are free to vary as we estimate the parameters of the model. The total sum of squares (TSS) is the sum of the squared difference between each PM
2.5 observation and the sample average, which is decomposed into two parts: the model sum of squares which is the part of the TSS that can be explained by the regression model, and the error sum of squares which is the remaining part of the TSS which cannot be explained by the model. The mean squares are the sum of squares divided by the respective degrees of freedom. The F value is the ratio of the model mean square to the error mean square, which is a classical statistical metric to quantify the explanatory power of the regression model and the final column is the
p-value associated with the F value which indicates the statistical significance of the model (lower means more significant). The F value and the
p-value seen in the table indicate that our model was able to explain a statistically significant portion of the data variation [
54,
55].
Table 3 is a parameter estimate table. The table indicates the significance of a parameter based on the t statistic used in the model. For our model, the table shows that AOD is the most significant variable in our model. This is to be expected, since AOD results from the presence of PM
2.5 in the atmosphere. However, AOD is not directly proportional to the mass concentration of PM
2.5, but is also dependent on the particle size distribution and other properties of the aerosol and the atmosphere. Relative humidity, temperature and ammonia are also significant variables in our model. The intercept is also relatively significant while pressure and wind speed are the least significant. We know this because Pr > F values have a significant range to them. Anything less than 0.01 has strong significance while anything between 0.01 to 0.05 has appreciable significances.
The model is dominated by data in the range of ~5 to ~25 micrograms per cubic meter. The model Pr < 0.0001 and F Value of ~39 (
Table 2), coupled with r
2 of ~0.5 (
Figure 6, Duplin County), suggest that the model is offering a good prediction overall. Moreover, having Pr for each of the variable (
Table 3) less than 0.05 suggests the robust contribution of each variable.
The results of the Duplin County prediction can be seen in
Figure 6. The figure shows four different scatter plots each representing a summer month (June, July and August) and a total graph that included all three months in one plot. These plots act to illustrate a visual representation of the model performance at this location.
Table 4 illustrates the normalized mean bias (NMB) and normalized mean error (NME) values, which are commonly used to assess the performance of models [
54,
55].
Once this model was verified for Duplin, it was then used to predict PM
2.5 concentrations in New Hanover and Catawba Counties. These locations are also affected by high agricultural production areas; however, they have very different meteorological conditions and processes affecting them daily. The results of this analysis can be seen in
Figure 7 and
Figure 8.
Figure 7 shows the results for New Hanover County and illustrates that despite the different meteorological mechanisms, the model can predict PM
2.5 concentrations for this area at a relatively high degree of accuracy.
Table 5 shows the model prediction parameters used for this study and clearly shows the model’s prediction struggle in July. This particularly high value is the result of the model’s dependence on AOD values. A high AOD value will sometimes result in an overpredicted PM
2.5 value.
Similarly,
Figure 8 shows the results for Catawba County; however, the model consistently under predicts concentrations at this location compared to both New Hanover and Duplin Counties, which can be seen in
Table 6. This indicates that the different meteorological mechanisms that dominate in Catawba County are not well understood by the model. Catawba County is also at a different elevation than the other two locations, which could explain some of the inconsistent predictions. Overall, these results suggest that the multiple regression model can predict (at a relatively high certainty) for the eastern portion of NC and loses some capabilities in the western portion of the state likely due to topography.
We then used the combination model to predict PM
2.5 values in Sampson County. As previously mentioned, Sampson County has one of the highest concentrations of agricultural activity (both animal and crop) in the state. Due to its rural landscape, there is no PM
2.5 monitor available and thus no PM
2.5 data for this location. Our model results showed, also seen in
Figure 9, that the PM
2.5 concentrations are low; however, the model did indicate that six days out of the nine-year period were over the EPA NAAQS of 35 μg m
−3. In order to further investigate how many exceedances were predicted by the model, the normalized mean error values calculated for New Hanover County were used to see how the model’s error would affect the number of exceedances calculated. New Hanover County errors were used because they were the highest of the three errors calculated for the different counties. Given the errors in the model, the number of exceedances could range from ten to three during the ten-year period. This could indicate that areas near agricultural activity could see higher PM
2.5 values, most of which are not being captured due to monitoring locations. This could also suggest that a reduction in ammonia emissions could have a positive impact on PM
2.5 concentrations in these high agricultural areas.
4. Conclusions
We have created a multiple regression model that can predict PM2.5 mass concentrations in counties with the highest agricultural activity in NC. High agricultural activity results in high concentration of ammonia being released from agricultural farms. Ammonia is also a precursor to PM2.5 formation and may cause an increase in PM2.5 concentrations in these areas which could lead to poor air quality and quality of life for those living in the area.
This model was developed using remotely sensed data (column ammonia and AOD) and ground-based meteorology, making it resistant to issues in the scarcity of ground-based PM
2.5 monitors around the state. Many of the areas around these agricultural farms are rural and have no monitor in place to track these concentrations daily. Satellite data, however, can introduce other limitations. For example, the instrument cannot directly measure ammonia concentrations at the surface, the concentrations are calculated through an algorithm developed by researchers which introduces limitation on the accuracy of the concentrations from the satellite. The same issues can be said for the AOD data. Despite the limitations, satellite algorithms are improving over time and as they improve, utilizing these data will become more important. Satellite data can also be utilized for the meteorology data if ground-based data are not available. Modern-era retrospective analysis for research and applications (MERRA) data is globally modeled meteorological data based on GEOS-5 atmospheric data assimilation system. The data models ground-based temperature, pressure and wind speed well. However, the data do not contain relative humidity, thus it must be obtained experimentally or mathematically. Moreover, PM
2.5 chemical composition analysis (
Figure 1) suggests that ammonium is a major component of the PM
2.5 aerosol.
The model was developed (i.e., training data) for the Cumberland and Johnston Counties. The model was validated (i.e., tested) for Duplin County and applied to predict PM2.5 in New Hanover, Catawba and Sampson Counties, all of which have a predominate influence from these agricultural farms. The model predicted PM2.5 values in Sampson County that were above the National Ambient Air Quality Standard (NAAQS) limit of 35 μg m−3. The need to investigate an ammonia reduction strategy due to its effects on PM2.5 concentrations in high agricultural areas is becoming more prevalent. If the PM2.5 reduction strategies seen in the past, such as the Clean Smokestacks Act and the Clean Air Act, have been as successful as they can be, then reducing ammonia emissions will not only provide air quality improvements, but also a reduction in PM2.5-related issues.