Quantifying the Influence of a Burn Event on Ammonia Concentrations Using a Machine-Learning Technique

Although combustion is considered a common source of ammonia (NH3) in the atmosphere, field measurements quantifying such emissions of NH3 are still lacking. In this study, online measurements of NH3 were performed by a cavity ring-down spectrometer, in the cold season at a rural site in Xianghe on the North China Plain. We found that the NH3 concentrations were mostly below 65 ppb during the study period. However, from 18 to 21 November 2017, a close burn event (~100 m) increased the NH3 concentrations to 145.6 ± 139.9 ppb. Using a machine-learning technique, we quantified that this burn event caused a significant increase in NH3 concentrations by 411%, compared with the scenario without the burn event. In addition, the ratio of ∆NH3/∆CO during the burn period was 0.016, which fell in the range of biomass burning. Future investigations are needed to evaluate the impacts of the NH3 combustion sources on air quality, ecosystems, and climate in the context of increasing burn events worldwide.


Introduction
As an important alkaline gas in the atmosphere, atmospheric ammonia (NH 3 ) has a crucial influence on atmospheric chemistry and the nitrogen cycle [1,2]. It can react with sulfuric acid (H 2 SO 4 ) and nitric acid (HNO 3 ), and enhance the formation of secondary inorganic aerosols (SIAs) [3]. Additionally, NH 3 can enhance the yield of secondary organic aerosols (SOAs) through aqueous chemistry [4]. These particulate SIAs and SOAs in the air decrease visibility, damage human health, and affect the climate [5][6][7][8][9]. After deposition, NH 3 can directly or indirectly affect terrestrial and aquatic ecosystems, such as soil acidification, water eutrophication, and reduction in biodiversity [10][11][12]. Thus, identifying and quantifying the sources of NH 3 is essential to understanding its vital role in atmospheric chemistry and reducing its negative impacts on the ecosystem and climate.
Although livestock waste and nitrogen fertilization are considered the most important sources of NH 3 emissions on a global or regional scale [13,14], NH 3 is also emitted into the atmosphere during the fuel combustion process through pyrolysis [15]. Biomass burning, such as forest fires, plays a critical role in NH 3 emissions in rural areas. For example, 1.4-8.2 and 0.7-2.6 Tg of NH 3 were emitted from forest fires in Indonesia in autumn in 2015, and Russia in July-August 2010, respectively [16,17], and 0.12 Mt of NH 3 was emitted from agricultural crop residue burning in India from 2008 to 2009 [18]. In urban areas, the nitrogen isotopic approach indicates that NH 3 originated primarily from combustion sources, including coal combustion, NH 3 slip from power plants, and vehicle exhausts [19][20][21]. A high-resolution global inventory revealed that NH 3 emissions from combustion sources continued to increase from 1960 to 2013 [22]. In China, an average of 8182 forest fires occurred from 1987 to 2007 [23]. The elevated fire frequencies were partially due to climate change [24]. Thus, the impact of combustion sources on NH 3 concentrations and related consequences needs further attention.
China is a global hotspot of atmospheric NH 3 emissions, with an annual increasing rate of 1.9% [25]. Notably, the North China Plain (NCP) is confirmed to be the largest region with high surface concentrations, and the highest emissions in China [26]. Despite the pollution reduction actions implemented since 2013, severe haze pollution events still occur in the cold season in this region. For example, the highest PM 2.5 concentration reached approximately 250 µg/m 3 in winter in 2020 [27], and it was dominated by sulfate-nitrateammonium (SNA), especially nitrate aerosols [28]. Recent studies have suggested that NH 3 plays an important role in determining nitrate concentrations [29,30]. Therefore, it is necessary to observe the concentrations of NH 3 and investigate its emission sources in this region.
In this study, we performed online measurements of NH 3 in the cold season in Xianghe, a rural site in NCP. During the observation campaign, we detected an unexpected burn event that had a significant influence on the NH 3 concentrations. Finally, we attempted to combine a novel machine-learning technique based on the random forest (RF) algorithm, to quantify the impact of burn events. Such an understanding could be beneficial for controlling NH 3 emissions and further improving air quality in the future.

Site Description
Since 2017, online measurements of NH 3 concentrations have been performed in Xianghe, NCP (39.75 • N, 116.96 • E) ( Figure 1). The site is surrounded by residential areas that lack tall buildings and obvious industrial emission sources. Based on its location in the northern part of the NCP between Beijing (~45 km) and Tianjin (~70 km), the instruments at the site can detect pollutants of urban, rural, background or mixed origins, reflecting the complex changes in NH 3 in the NCP.
into the atmosphere during the fuel combustion process through pyrolysis [15]. Biomass burning, such as forest fires, plays a critical role in NH3 emissions in rural areas. For example, 1.4-8.2 and 0.7-2.6 Tg of NH3 were emitted from forest fires in Indonesia in autumn in 2015, and Russia in July-August 2010, respectively [16,17], and 0.12 Mt of NH3 was emitted from agricultural crop residue burning in India from 2008 to 2009 [18]. In urban areas, the nitrogen isotopic approach indicates that NH3 originated primarily from combustion sources, including coal combustion, NH3 slip from power plants, and vehicle exhausts [19][20][21]. A high-resolution global inventory revealed that NH3 emissions from combustion sources continued to increase from 1960 to 2013 [22]. In China, an average of 8182 forest fires occurred from 1987 to 2007 [23]. The elevated fire frequencies were partially due to climate change [24]. Thus, the impact of combustion sources on NH3 concentrations and related consequences needs further attention.
China is a global hotspot of atmospheric NH3 emissions, with an annual increasing rate of 1.9% [25]. Notably, the North China Plain (NCP) is confirmed to be the largest region with high surface concentrations, and the highest emissions in China [26]. Despite the pollution reduction actions implemented since 2013, severe haze pollution events stil occur in the cold season in this region. For example, the highest PM2.5 concentration reached approximately 250 μg/m 3 in winter in 2020 [27], and it was dominated by sulfatenitrate-ammonium (SNA), especially nitrate aerosols [28]. Recent studies have suggested that NH3 plays an important role in determining nitrate concentrations [29,30]. Therefore it is necessary to observe the concentrations of NH3 and investigate its emission sources in this region.
In this study, we performed online measurements of NH3 in the cold season in Xianghe, a rural site in NCP. During the observation campaign, we detected an unexpected burn event that had a significant influence on the NH3 concentrations. Finally, we attempted to combine a novel machine-learning technique based on the random forest (RF) algorithm, to quantify the impact of burn events. Such an understanding could be beneficial for controlling NH3 emissions and further improving air quality in the future.

Site Description
Since 2017, online measurements of NH3 concentrations have been performed in Xianghe, NCP (39.75° N, 116.96° E) ( Figure 1). The site is surrounded by residential areas that lack tall buildings and obvious industrial emission sources. Based on its location in the northern part of the NCP between Beijing (~45 km) and Tianjin (~70 km), the instruments at the site can detect pollutants of urban, rural, background or mixed origins, reflecting the complex changes in NH3 in the NCP.   Hourly concentrations of NH 3 were measured at a high temporal resolution of 1 Hz, online, using a standard cavity ring-down spectrometer (CRDS) (G2103, Picarro Inc., Santa Clara, CA, USA). CRDS is a direct absorption technique that uses pulsed or continuous light sources and has a significantly higher sensitivity than conventional spectrometers [31]. In CRDS, two high reflectivity mirrors are used in the optical cavity to increase the absorption optical path length, thereby enhancing the contrast of the absorption signal of NH 3 . The CRDS setting measures how long it takes for the light to drop to a certain percentage of its original intensity, and the "ring-down time" is used to calculate the concentration of NH 3 in the cavity [32].
To prevent water vapor from affecting the NH 3 spectrum, the manufacturer has incorporated a correction procedure for the reported NH 3 values [33]. In addition, to reduce the adsorption of NH 3 , the Teflon tubing was insulated and warmed with heating tape (~45.7 • C). Meanwhile, a filter was installed at the front of the inlet to induce ambient air flow. According to the air conditions, the filter was replaced every 2 weeks to 1 month. The instrument was placed in an air-conditioned cabin laboratory; more detailed descriptions are documented elsewhere [34].

Other Supporting Data
Air pollutants and meteorological data are also used in this study. The hourly concentrations of SO 2 , NO x , CO, and PM 2.5 were measured at the same height as NH 3 . Meteorological parameters, including temperature (T), relative humidity (RH), wind speed (WS), and wind direction (WD), were obtained from the China Meteorological Administration (http://data.cma.cn (accessed on 14 October 2021)).

Burn Event
During the observation campaign, a burn event occurred from 18 to 21 November, 2017. This burn event was caused by the combustion of garbage in the nearby residential area, which was approximately 100 m away from the observation site. The burned material was complex and included discarded paper, kitchen waste, crop residues, weeds, branches, and leaves.

RF Models
The temporal variations in NH 3 in this study demonstrated a cycle ranging from 4 to 7 days. These strong cycles are regional in nature and controlled by the passage of cold fronts [35]. This unique temporal feature indicated that the results presented are reproducible on a regular basis, and applicable for the RF model in predicting NH 3 dynamics with meteorological conditions.
In this study, we used a machine-learning technique to quantify the influence of burn events on NH 3 concentrations. First, we established a model (RF1) with observed NH 3 concentrations as the dependent variable, and predictors (meteorological parameters, time predictors, and regional transport parameters) as the independent variables (Table 1). RF1 was trained on datasets during the nonburning period (8-17 and 22-30 November 2017). The training set accounted for 80% of data, and the testing set included the remaining 20%. On the basis of RF1, a series of models (RF2-RF11) were established, according to the relative importance of the predictor variables for eliminating or adding some predictors (Table S1). A detailed evaluation of all the established models is provided in Table S2, including the coefficients of determination (R 2 ), the fraction of predictions within a factor of 2 (FAC2), mean bias, normalized mean bias, root-mean-square error (RMSE), and Pearson correlation coefficient (PCC). The performance of RF5 was considered the best due to its higher R 2 and lower RMSE.
The RF models were developed using the rmweather R package. However, unlike previous studies, we further adjusted two important parameters of RF5, namely, the number of trees (ntree) and the number of variables split in each node (mtry). Mtry values from 1 to 8 with an interval of 1, and ntree values from 50 to 500 with an interval of 50, were selected. As shown in Figure S1, when ntree was 300 and mtry was 5, the simulation effect of the model was best, as indicated by the higher R 2 value and lower prediction error (MSE). After the RF model was adjusted and optimized, the testing datasets were randomly selected to assess the correlation between the observed and predicted concentrations ( Figure S2), to ensure that this model could make better predictions. Finally, we used this model to predict NH 3 concentrations under the assumption of no burn events. To evaluate the impact of the burn events (18-21 November 2017) on NH 3 concentrations, relative changes (R) between the observed concentrations (C observed ) and predicted concentrations (C predicted ) for NH 3 were defined using the following equation:

Changes in Observed Concentrations of NH 3
Hourly concentrations of NH 3 were selected from 8 to 30 November 2017 in this study, and their temporal variations are shown in Figure 2. We found that the NH 3 concentrations were mostly below 65 ppb during the study period, with the exception of 18-21 November, when a burn event occurred. Therefore, in the following sections, we have separated the dataset into two periods, i.e., the burning period (18-21 November 2017) and the nonburning period (8-17 and 22-30 November 2017).
During the nonburning period, the NH 3 concentrations in Xianghe ranged from 1.9 to 154.3 ppb, with a mean value of 25.4 ± 16.9 ppb. The NH 3 concentrations in this study were comparable to the urban observations (28.5 ± 11.6 ppb) in autumn in Beijing [36]. Moreover, the average NH 3 level measured in Xianghe was generally similar to observations in India (24.6 ± 5.0 ppb) [37], but much higher than those in Europe (1.2 ppb) and the United States (2.4 ppb) in autumn [10,38].
After the start of the burn event, the observed concentrations of NH 3 significantly increased, with the highest value exceeding 600 ppb. In addition, three concentration spikes occurred: at 7:00, 18 November (281.6 ppb); 2:00, 19 November (547.1 ppb); and 10:00, 20 November (601.4 ppb). During the burning period, the average NH 3 concentration was 145.6 ± 139.9 ppb, which is five times that in a previous report (26.6 ± 13.9 ppb) at the Xianghe site in the cold season [39]. During the nonburning period, the NH3 concentrations in Xianghe ranged from 1.9 to 154.3 ppb, with a mean value of 25.4 ± 16.9 ppb. The NH3 concentrations in this study were comparable to the urban observations (28.5 ± 11.6 ppb) in autumn in Beijing [36]. Moreover, the average NH3 level measured in Xianghe was generally similar to observations in India (24.6 ± 5.0 ppb) [37], but much higher than those in Europe (1.2 ppb) and the United States (2.4 ppb) in autumn [10,38].
After the start of the burn event, the observed concentrations of NH3 significantly increased, with the highest value exceeding 600 ppb. In addition, three concentration spikes occurred: at 7:00, 18 November (281.6 ppb); 2:00, 19 November (547.1 ppb); and 10:00, 20 November (601.4 ppb). During the burning period, the average NH3 concentration was 145.6 ± 139.9 ppb, which is five times that in a previous report (26.6 ± 13.9 ppb) at the Xianghe site in the cold season [39].
In addition to NH3, the concentrations of other air pollutants also significantly increased during the burning period, e.g., by 209.7%, 84.3%, 69.1%, and 79.8% for PM2.5, SO2, NOx, and CO, respectively ( Figure S3). These air pollutants had a positive correlation with NH3 during the burning period, especially CO (R 2 = 0.74, p < 0.01) and NOx (R 2 = 0.72, p < 0.01). However, their positive correlation was not significant in the nonburning period (Table S3). These results all indicated the potential influence of combustion sources on air quality. In addition, the intensity of combustion was closely related to weather conditions [40]. In our study, the NH3 concentrations were higher during the burning period when the air mass originated from the northwest, and had a higher RH (56.2 ± 26.3%) and lower T (−1.1 ± 5 °C) and WS (1.6 ± 1.5 m s −1 ) ( Figure S4). To quantify the impact of the burn events on NH3, we predicted NH3 concentrations without burn events using a machinelearning technique in the next section.

Changes in Predicted Concentrations of NH3
We first established a model (RF5) to predict the NH3 concentrations assuming that no burn event occurred from 18 to 21 November 2017. The predicted results are shown in Figure 2, with all predicted concentrations lower than 65 ppb. In addition, the temporal variation in predicted NH3 concentrations was basically similar to its observed value in the nonburning period. This finding indicates that the concentrations and temporal pattern of NH3 will not change significantly if there is no burn event. In addition to NH 3 , the concentrations of other air pollutants also significantly increased during the burning period, e.g., by 209.7%, 84.3%, 69.1%, and 79.8% for PM 2.5 , SO 2 , NO x , and CO, respectively ( Figure S3). These air pollutants had a positive correlation with NH 3 during the burning period, especially CO (R 2 = 0.74, p < 0.01) and NO x (R 2 = 0.72, p < 0.01). However, their positive correlation was not significant in the nonburning period (Table S3). These results all indicated the potential influence of combustion sources on air quality. In addition, the intensity of combustion was closely related to weather conditions [40]. In our study, the NH 3 concentrations were higher during the burning period when the air mass originated from the northwest, and had a higher RH (56.2 ± 26.3%) and lower T (−1.1 ± 5 • C) and WS (1.6 ± 1.5 m s −1 ) ( Figure S4). To quantify the impact of the burn events on NH 3 , we predicted NH 3 concentrations without burn events using a machine-learning technique in the next section.

Changes in Predicted Concentrations of NH 3
We first established a model (RF5) to predict the NH 3 concentrations assuming that no burn event occurred from 18 to 21 November 2017. The predicted results are shown in Figure 2, with all predicted concentrations lower than 65 ppb. In addition, the temporal variation in predicted NH 3 concentrations was basically similar to its observed value in the nonburning period. This finding indicates that the concentrations and temporal pattern of NH 3 will not change significantly if there is no burn event.
The difference between the predicted and observed concentrations of NH 3 allows us to quantify the impacts of burn events on NH 3 emissions. As shown in Figure 3 and Table S4, remarkable differences were not found between the observed and predicted concentrations of NH 3 before the burn event, indicating that our model captured the variations in NH 3 concentrations well. However, the observed concentration of NH 3 increased drastically compared with the predicted value during the burning period, with increased ratios of 336.9%, 561.1%, and 652.2% in the first three days of the burn event. At the end of the burning period and with the resumption of "normal emissions", the predicted values were consistent with the observations after 22 November. Overall, the observed and predicted concentrations of NH 3 during the burning period were 145.6 ± 139.9 and 28.5 ± 10.4 ppb, Atmosphere 2022, 13, 170 6 of 9 respectively, resulting in an increase of 411%. These results indicated that the combustion source is an important factor of NH 3 emissions.
S4, remarkable differences were not found between the observed and predicted concentrations of NH3 before the burn event, indicating that our model captured the variations in NH3 concentrations well. However, the observed concentration of NH3 increased drastically compared with the predicted value during the burning period, with increased ratios of 336.9%, 561.1%, and 652.2% in the first three days of the burn event. At the end of the burning period and with the resumption of "normal emissions", the predicted values were consistent with the observations after 22 November. Overall, the observed and predicted concentrations of NH3 during the burning period were 145.6 ± 139.9 and 28.5 ± 10.4 ppb, respectively, resulting in an increase of 411%. These results indicated that the combustion source is an important factor of NH3 emissions.

Dominant Source of NH3 during the Burn Event
Although the overall impacts of the burn event on NH3 concentrations were identified in the above sections, the potential dominant source contributing to the substantial increase in NH3 was still unclear. To address this concern, we further investigated the relationship between NH3 and other air pollutants.
As shown in Figure 4, we found that the ratio of NH3 to other air pollutants significantly increased after the burn event occurred. Compared with those before burn events, the ratio of NH3/PM2.5, NH3/SO2, NH3/NOx, and NH3/CO increased by a factor of 1.6, 2.5, 3.0, and 5.4 during the burning period, respectively (Table S5). This finding indicated a dramatic change of CO during the burn event. Thus, CO was selected as a tracer of fire [41]. Then, the emission ratio (ER) was defined as ∆NH3/∆CO in our study [17,42], where ∆NH3 and ∆CO represented the difference in the corresponding concentrations of NH3 and CO before and during the burn event. In this study, the calculated ∆NH3/∆CO value was 0.016, which was consistent with the characteristics of biomass burning (Table 2). Therefore, the dominant source emitting substantial NH3 might be biomass burning in this burn event. Future control measures on NH3 emissions should pay more attention to potential contributions from biomass burning.

Dominant Source of NH 3 during the Burn Event
Although the overall impacts of the burn event on NH 3 concentrations were identified in the above sections, the potential dominant source contributing to the substantial increase in NH 3 was still unclear. To address this concern, we further investigated the relationship between NH 3 and other air pollutants.
As shown in Figure 4, we found that the ratio of NH 3 to other air pollutants significantly increased after the burn event occurred. Compared with those before burn events, the ratio of NH 3 /PM 2.5 , NH 3 /SO 2 , NH 3 /NO x , and NH 3 /CO increased by a factor of 1.6, 2.5, 3.0, and 5.4 during the burning period, respectively (Table S5). This finding indicated a dramatic change of CO during the burn event. Thus, CO was selected as a tracer of fire [41]. Then, the emission ratio (ER) was defined as ∆NH 3 /∆CO in our study [17,42], where ∆NH 3 and ∆CO represented the difference in the corresponding concentrations of NH 3 and CO before and during the burn event. In this study, the calculated ∆NH 3 /∆CO value was 0.016, which was consistent with the characteristics of biomass burning (Table 2). Therefore, the dominant source emitting substantial NH 3 might be biomass burning in this burn event. Future control measures on NH 3 emissions should pay more attention to potential contributions from biomass burning.

Conclusions
In this study, the NH3 concentrations were measured by the CRDS technique in the cold season at a rural site in the NCP. The hourly NH3 concentrations were mostly below 65 ppb during the study period. However, an unexpected burn event caused a significant increase in NH3 concentration from 18 to 21 November 2017, with peak values exceeding 600 ppb. With the aid of a machine-learning technique, we found that the burn event could cause a 411% increase in the NH3 concentration. Notably, the ∆NH3/∆CO ratio was 0.016 during the burning period, indicating that biomass burning might be the dominant emission source. Due to the increasing occurrence of burn events worldwide, the impacts of the combustion sources of NH3 on air quality, ecosystems, and climate need to be further explored.
Supplementary Materials: The following are available online at www.mdpi.com/xxx/s1: Figure S1: Influence of different ntree and mtry values on RF model; Figure S2: Performance of the RF model in predicting the NH3 concentrations; Figure S3: Temporal variation of hourly concentrations of

Conclusions
In this study, the NH 3 concentrations were measured by the CRDS technique in the cold season at a rural site in the NCP. The hourly NH 3 concentrations were mostly below 65 ppb during the study period. However, an unexpected burn event caused a significant increase in NH 3 concentration from 18 to 21 November 2017, with peak values exceeding 600 ppb. With the aid of a machine-learning technique, we found that the burn event could cause a 411% increase in the NH 3 concentration. Notably, the ∆NH 3 /∆CO ratio was 0.016 during the burning period, indicating that biomass burning might be the dominant emission source. Due to the increasing occurrence of burn events worldwide, the impacts of the combustion sources of NH 3 on air quality, ecosystems, and climate need to be further explored.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/ 10.3390/atmos13020170/s1: Figure S1: Influence of different ntree and mtry values on RF model; Figure S2: Performance of the RF model in predicting the NH 3 concentrations; Figure S3: Temporal variation of hourly concentrations of PM 2.5 , SO 2 , NO x , and CO; Figure S4: Hourly variation of meteorological parameters; Table S1: Summary of 11 RF models and their predictors in this study; Table S2: Model validation in 11 RF models in this study; Table S3: Correlation matrix of NH 3 , other air pollutants and meteorological parameters during the burning period; Table S4: Observed and predicted concentrations of NH 3 at the Xianghe site during the burning period; Table S5: Range and average of the ratio between NH 3 and other air pollutants before and during the burn event.