Impact on air quality and health due to the Saddleworth Moor fire in northern England

This study uses WRF-Chem v3.7.1 (Grell et al 2005), a fully coupled atmospheric chemistry model at 10 km resolution, to simulate hourly PM_2.5 concentrations during the Saddleworth Moor fires 2018. The study domain covers northern England (−4.9–0.7°E and 53.0–54.4°N) and a population of 14 million people. Our simulations are performed using the same model version and set-up as Conibear et al (2018a), Reddington et al (2019) and Kiely et al (2019). For a more detailed model description refer to Conibear et al (2018a).

Meteorological initial boundary conditions (IBC) were provided by the National Centers for Environmental Prediction (NCEP) Global Forecasting System (GFS) reanalysis (meteorology) at 6-hour time steps and 0.5° resolution. Chemical IBC are from the Whole Atmosphere Community Climate Model (WACCM) (Marsh et al 2013) (NCAR 2018). An updated MOZART-4 (Emmons et al 2010) scheme is used to calculate gas-phase chemical reactions. Aerosol dynamics and processes were represented by the Model for Simulating Aerosol Interactions and Chemistry (MOSAIC), which included aqueous chemistry and extended treatment of organic aerosol (Hodzic and Knote 2014). Four bins were used represent aerosol size: 0.039–0.156 µm, 0.156–0.625 µm, 0.625–2.5 µm and 2.5–10 µm. WRF was nudged on all 33-vertical terrain following levels every 3 h in order to keep mesoscale meteorology in line with the reanalysis meteorology from GFS (NCEP 2007). Variables nudged included horizontal and vertical wind, potential temperature and water vapour mixing ratio.

Monthly anthropogenic emissions were from the Emission Database for Global Atmospheric Research with Task Force on Hemispheric Transport of Air Pollution version 2.2 (EDGAR-HTAP2) (Janssens-Maenhout et al 2015) at 0.1° resolution for 2010 (see SM section 3.2 for more information). Biogenic emissions were calculated online by the Model of Emissions of Gases and Aerosols from Nature (MEGAN) (Guenther et al 2006). Dust emissions are also calculated online using the GOCART with Air Force Weather Agency (AFWA) modifications (LeGrand et al 2019).

We calculate the contribution of the fires between June 16th and July 14th 2018 on PM_2.5 surface concentrations by comparing simulations with and without fire emissions included.

Wildfire emissions are taken from the Fire Inventory from NCAR version 1.5 (FINNv1.5). The FINNv1.5 emissions dataset combines satellite observations, land cover, biomass consumption estimates and emissions factors to calculate fire emissions globally at 1 km resolution every day. Satellite observations from the MODIS Thermal Anomalies Product provide detections of active fires with a nominal horizontal resolution of ∼1 km². Burned area is assumed to be 1 km² for each fire identified and scaled back based on the density of vegetation from the MODIS Continuous Fields (VCF) (i.e. if 50% bare =0.5 km² burned area). The type of vegetation burned during a detected fire is determined using the MODIS Collection 5 Land Cover Type (LCT). This assigns each fire pixel to one of 16 possible land cover/land use classes and also the density of vegetation at 500 m resolution, scaled to 1 km. The 16 land cover types are then aggregated into 8 generic categories to which fuel loadings are applied (Wiedinmyer et al 2011). Fuel loadings are from Hoelzemann et al (2004) and emissions factors are from Akagi et al (2011), McMeeking (2008) and Andrae and Merlet (2001). FINNv1.5 includes all emissions from above ground vegetation but not from the combustion of peat (Kiely et al 2019).

We compare the FINN burned area (1 km resolution) with MODIS burned area (500 m resolution) in order to evaluate whether the resolution of the MODIS hotspot data used within FINN to estimate emissions is able to represent the fire size correctly, and thus emissions. We find the burned areas to be very similar for FINN (9.77 km²) and MODIS (8.43 km²) and the datasets to be in agreement spatially. The burned area in FINN is likely to be slightly higher than MODIS because of the lower resolution of the dataset. Since the Saddleworth Moor fires occurred on an area that is dominated by peat bog, with overlying vegetation including heather, grass and juniper (GMCA; Xu et al 2018) (figure S1), we are confident the need for scaling emissions is due to the missing peat emissions rather than an error relating to fire size.

We therefore scale all FINN emissions over the Saddleworth Moor region (figure 1 and figure S1) to account for the underestimation of emissions in the dataset due to the missing peat emissions. Scaling is performed equally across all FINN emission species, and is altered daily to match the daily mean observations of PM_2.5 at AURN sites (see table S2 for AURN sites). We scale emissions by a factor 5 on June 26th and a factor 10 on 27th, 28th and 29th. On all other days, we use the original unscaled FINN emissions. More details on the evaluation of FINN scaling can be found in the supplementary material (SM sections 2 and 4 and figures S2 and 3).

Hourly observations of PM_2.5 from the Automated Rural and Urban Network (AURN) in the UK are used to evaluate the model's performance at hourly and daily temporal resolution. We evaluate the model against all AURN sites in the north-west and Yorkshire and Humber regions of England, which are mostly urban sites (see table S1 and figure 1 for more details). Daily means are calculated from hourly data for days where >90% of data is available at a given site.

The health impact from short-term exposure to elevated pollutants from the Saddleworth Moor fires can be calculated using an exposure response function:

$$\begin{equation}{E_m} = \,\mathop \sum \limits_{i = 1}^N {B_d}.po{p_i}.A{F_i}\,\end{equation} \tag{ 1 }$$

where:

$$\begin{equation}AF = \left( {\frac{{RR - 1}}{{RR}}} \right)\end{equation} \tag{ 2 }$$

and:

$$\begin{equation}RR = ex{p^{\beta \left( {X - {X_0}} \right)\,}}\end{equation} \tag{ 3 }$$

E_m represents the excess mortality caused by exposure to PM_2.5 over the safe limit of exposure (X–X₀) each day. N is the number of days within the simulation and i is the day in simulation, B_d is the baseline death rate, pop is the population exposed and AF is the attributable fraction of mortality due to exposure to PM_2.5. The AF is calculated using the concept of relative risk (RR), this is the probability of mortality from a disease endpoint within an exposed population compared to the probability of mortality within an unexposed population (ß). The concentration a population is exposed to is given by X and the safe-limit of exposure is X₀. Since there is little evidence to suggest a safe-limit of exposure to PM_2.5 we assume X₀ to be zero (Holgate 1998, Schmidt et al 2011, Macintyre et al 2016). We use beta values from Atkinson et al (2014) for PM_2.5 (1.04% (95% CI: 0.52%, 1.56%) per 10 μg m⁻³ increase). Since short-term health impacts are assumed to be equal across ages, we use baseline mortality rates for all ages and population for all ages in the calculations (Atkinson et al 2014).

The health impact assessment is carried out using the 'subtraction' method, which is the one most commonly used in short-term health impact studies (Macintyre et al 2016; Crippa et al 2016). This method calculates the excess mortality from fires (E_{m FIRES ONLY}) to be the difference between the excess mortality from PM_2.5 in the simulation with fires on (E_{m FIRES ON}) and excess mortality from PM_2.5 when there are no fires in the simulation (E_{m FIRES OFF}) (equation (4)).

$$\begin{equation}{E_{m\,FIRES\,ONLY\,}} = \,{E_{m\,FIRES\,ON}} - {E_{m\,FIRES\,OFF}}\end{equation} \tag{ 4 }$$

We use population count data from the Gridded Population of the World, Version 4.11 (CIESIN 2018) for 2015 at 5 km resolution. The dataset is created by the Centre for International Earth Science Information Network (CIESIN) and was accessed from the National Aeronautics and Space Administration (NASA) Socioeconomic Data and Applications Centre (SEDAC). The dataset uses estimates of human population based on the national census and population registers. Input data from 2005–2014 are extrapolated to produce estimates of population for 5-year increments. A map of this population data in the north-west of England is available for reference (figure 1). Baseline mortality rate data for north-west England is taken from the Global Burden of Disease for 2015 (IMHE 2018).

The economic cost of mortality caused by exposure to PM_2.5 from the fires is calculated using the 'Value of Prevented Fatality' (VPF) from the UK Department for Transport. The VPF was initially used to evaluate transport projects which entail expected reductions in fatalities (Guria et al 2005). However, in recent years, many other government sectors, such as the UK Environment Agency and Health Protection Agency, have begun to utilise the concept (Deloitte 2009). Several other studies have used this method to quantify the benefits of air quality improvements from reduced mortality (Krupnick et al 1996, US EPA 1997). There is a large range in estimates for VPF so we use values from the Department for Transport since these are based on UK costs of mortality and lie within the range of other estimates (see SM table S4 and SM section 6 for more details). The estimates are broken down into human costs, medical costs, lost output and other costs. Values are given in GBP for 2008, which we scale to 2018 values in line with inflation (Bank of England 2019). The human cost component reflects the pain and suffering felt by the victim and relatives and the reduction in life quality during the period of injury. Thus, the human cost is derived from the 'Willingness to Pay' of the population to reduce this risk. Medical costs represent all treatment costs and lost output represents working days lost and therefore, the total expected lost earnings before tax, as well as national insurance payments. Other costs represent emergency services and benefits.

AURN observations (figure 2) indicate PM_2.5 concentrations at 5 locations in the north-west exceeded 100 µg m⁻³ between June 16th and July 14th 2018. As expected, concentrations were highest at sites near to Saddleworth Moor, reaching 140 and 225 µg m⁻³ on June 27th and 28th at Manchester Piccadilly and Salford Eccles respectively (figure 2). They also reached >175 µg m⁻³ at Wigan Centre, 50 km from the fires. Other sites in the network were relatively unaffected by the fires, with little variation in concentrations during the fires (figure 2).

Zoom In Zoom Out Reset image size

The model is evaluated using Pearson correlation coefficient (r), mean bias (MB), normalised mean bias (NMB), root mean square error (RMSE), mean absolute error (MAE) and normalised mean absolute error (NMAE). Firstly, when assessing model performance at the daily resolution, without scaling FINN fire emissions (no scaling), the model performs relatively poorly (table 1 and figure S2). The Pearson correlation score is similar to the simulation without fire emissions (no fires) (0.69 and 0.62 respectively). RMSE (5.06 µg m⁻³), NMB (−0.22) and NMAE (0.33) are also similar to with no fire emissions (5.30 µg m⁻³, −0.24 and 0.34 respectively). Referring to the time series from each AURN site it becomes clear the poor performance of the model is dominated by sites where fire emissions are not being captured well (see figures S2 and S3). When a factor 10 scaling (10 × scaling) is applied to FINN emissions over Saddleworth Moor the correlation is improved (0.74) and NMB is substantially improved (0.10) (table 1). The RMSE (4.31 µg m⁻³) and NMAE (0.31) also improve. However, the model still over predicts PM_2.5 in the early stages of the fire (see SM figure 2). The over prediction may be due to a change in fuel source through the fire lifetime, from the surface vegetation (heather and grass) initially, which FINN accounts for, to underlying peat once the surface vegetation has been consumed (GMCA 2019). Peat has much higher emissions per unit burnt (9.1 g kg⁻¹ burned (but estimates range from 6–30 g kg⁻¹) compared with 6.3–15.3 g kg⁻¹ for other vegetation types burned (GFEDv4)). Alongside this, FINNv1.5 does not account for whether burning is smouldering or flaming, which can change emissions significantly, particularly in peat fires where emissions are highest during smouldering due to colder combustion temperatures (Stockwell et al 2016). To try to account for the change in fuel type we perform a simulation where we adjust the scaling of the fire emissions each day (see section 2.2 for more details on scaling). Using a daily time-variant scaling (Time-varying scaling) we improve RMSE (4.27 µg m⁻³) and Pearson correlation (0.77) and find the simulation has a similar NMAE (0.31) when compared with factor 10 scaling (10 × scaling: RMSE = 4.31 µg m⁻³, r = 0.74 and NMAE = 0.31) (see table 1). Time-varying scaling also performs best at hourly time resolution with improved correlation (0.42), RMSE (7.11 µg m⁻³) and NMAE to 0.47 (compared with factor 10 scaling and no scaling simulations (r = 0.37 and 0.351, RMSE = 7.557 and 7.78 µg m⁻³))and the removal of the over prediction at the start of the fires (see SM figure 2). We therefore use the time-variant scaling of FINN emissions as our best-estimate (more details in SM sections 2 and 4)

Using WRF-Chem simulations we calculate the percentage increase in PM_2.5 at the surface due to fires as $\left( {\frac{{P{M_{2.5\,Fires}} - P{M_{2.5\,No\,Fires}}}}{{P{M_{2.5\,No\,Fires}}}} \times \;100} \right)$, where PM_2.5 from the simulation with fire emissions is labelled as PM_{2.5 Fires} and PM_2.5 from the simulation without fire emissions is labelled as PM_{2.5 No Fires}. Between June 23rd and June 30th 2018 model simulations indicate the mean increase in PM_2.5 due to fires (figure 3(c)) is largest in the area surrounding Oldham (>300% increase). However, there is also a 150%–200% increase in PM_2.5 in Manchester, Bolton and Wigan. Areas as far away as Liverpool, Preston and Warrington are also affected with 10%–50% increases in PM_2.5 observed. Daily mean percentage increase in PM_2.5 from fires indicates that the largest increase in PM_2.5 observed is due to the Saddleworth Moor fires (figure S6) on the 26th and 27th June. Results indicate PM_2.5 increases of >600% in Manchester, Bolton and Wigan and >1000% in Oldham were due to the fires (figure S6). Large areas of the north-west also experience >350% increase in PM_2.5, including Wigan, 50 km from the fires, and a 100% increase is observed as far west as the Irish Sea. Simulations indicate the Winter Hill fires on June 29th and 30th were also associated with PM_2.5 increases of 100 to >600% in Bolton, Wigan and Southport (40 km away) (figure S6). The Winter Hill fire was substantially smaller and occurred further north where the population density is lower. In summary, WRF-Chem simulations of wildfire impacts on atmospheric composition indicate an extensive area in which particulate matter concentrations were enhanced, far above normal regional and UK levels.

Zoom In Zoom Out Reset image size

To put these results into the context of air quality guidelines, we use the Daily Air Quality Index (DAQI) values and the World Health Organisation (WHO) 24-hour guideline for PM_2.5 combined with population count to estimate the population exposure (figures 3(a) &(b)). A limitation of this method is that it assumes the population living in the affected area is exposed to PM_2.5 from the fires, however this may vary based on whether the environment they are in provides any passive or active filtration (e.g. indoor air filtration) and does not account for how much time is spent outdoors. The DAQI is used to advise the UK population on recommended behaviour changes during air pollution events. For example, the advice for PM_2.5 within the very high DAQI band is for everyone to reduce outdoor activity, and for those with asthma to be aware for the potential need for increased medication (see table S2 for more details on DAQI bands). Between June 23rd and June 30th 0.8 million people were exposed to the highest DAQI band (very high: >71 µg m⁻³) in areas close to the Saddleworth Moor fire (figure 3(a)). This exposure was dominated by PM_2.5 on June 27th (0.5 million exposed) but 0.2 million people were also exposed to very high levels of PM_2.5 on June 26th (see figure S4) (note totals may not add up due to rounding). 0.8 million people were exposed to concentrations above 54 µg m⁻³ (high DAQI: 54–70 µg m⁻³) and 1.3 million people to 36–54 µg m⁻³ (moderate DAQI) (figure 3(a)). The degradation in air quality was dominated by the Saddleworth Moor fires since exposure to the Winter Hill fire accounted for only 5% of the total moderate DAQI exposure (0.06 million to moderate DAQI levels on June 30th) (see figure S4). This is likely in part because the area surrounding Winter Hill is more sparsely populated. Nonetheless, these results indicate almost a quarter of the population within our simulation domain (22% of the total 14 million people in the model domain) were exposed to concentrations of >36 µg m⁻³ on at least one day between June 23rd and 30th due to the Saddleworth Moor and Winter Hill fires. When we compare these results with the PM_{2.5 No Fires} simulation (figure S7), in which no day exceeds the low DAQI, it is clear that the fires are responsible for the degradation in air quality during this time period.

We also frame our results in the context of the WHO 24-hour guideline of 25 µg m⁻³ (figure 3(b)). Results show that 4.5 million people were exposed to PM_2.5 above this guideline for at least one 24-hour period between June 23rd and 30th. The impact was widespread, affecting Oldham, Manchester, Wigan and areas of high population on the coast north of Liverpool (for further detail see figure S5).

In addition, we examine the fraction of the total annual DAQI high (48–71) and very high (71+) hourly exceedances which the fires represent at each of the AURN sites used in this study. We find that hourly DAQI exceedances during the fire period (June 23rd–June 30th) represented a large fraction of the total annual high (48–71) and very high (71+) DAQI hourly exceedances at many sites (table S5(a)). At Manchester Piccadilly, Salford Eccles and Wirral Tranmere 31%, 77% and 58% of total annual hourly DAQI very high exceedances occurred within the week of the fires (see table S5(b) for more details). Thus, not only did the fires have a large impact on air quality between June 23rd and 30th but they also represented a large fraction of the annual hourly DAQI exceedances.

Finally, we calculate the short-term mortality burden due to exposure to PM_2.5 from the fires and the economic cost, using the subtraction method detailed in section 2.5. We use the concentration response function of Atkinson et al (2014) with no assumed safe-limit of exposure (0 µg m⁻³) since there is little evidence in the current literature to suggest that there is a safe level. In total over the 7-day period of the fires there were 28 (95% CI: 14.1–42.1) deaths brought forward with a mean daily excess mortality of 3.53 deaths per day (95% CI: 1.77–5.26) (figure 4(a)). This comprises a large fraction of the total 81 deaths brought forward over the four-week simulation (16th June–14th July). When the fraction of daily mortality from fires is calculated (i.e. $\frac{{{{\text{E}}_{{\text{m Fires}}}} - {{\text{E}}_{{\text{m No Fires}}}}}}{{{{\text{E}}_{{\text{m Fires}}}}}} \times 100$) the impact of the fires is even more apparent (figure 4(a)). On June 23rd and 24th, the fraction of total excess mortality caused by fires is very low (0.08%–0.9%) but substantially increases during the fires (25th and 26th–11%–39%), peaking at ∼60% on June 27th and 33% on June 30th. Thus, the increase in excess mortality observed during June 23rd–30th 2018 was dominated by the Saddleworth Moor and Winter Hill fires

Zoom In Zoom Out Reset image size

In order to make our results comparable to other research in the literature we also calculate the percentage increase in excess mortality (E_m) due to short-term exposure to PM_2.5 from the fires only ($\frac{{{E_{m Fires}} - {E_{m No Fires}} }}{{{E_{m No Fires}}}} \times 100)$. This gives a result that is independent of the population size, since the population in the domain is relatively small (14 million) in comparison to other studies. The results indicate that up to 3.8 of 6.4 excess mortalities were due to exposure to PM_2.5 from the fires, representing a 165% (95% CI: 84%–246%) increase in E_m across the region due to exposure to PM_2.5 from the fires (figure 3(b)). While, the Winter Hill fire was associated with 1.9 of 2.8 total excess mortalities, a 96% (95% CI: 48%–131%) increase in E_m. In total over the 7-day period of the fires there were 7 (95% CI: 4–11) deaths brought forward with a mean daily excess mortality of 0.9 deaths per day (95% CI: 0.5–1.4) (figure 4(a)). This comprises a large fraction of the total 20 deaths brought forward over the four-week period simulated (June 16th–July 14th 2018). Thus, the increase in excess mortality observed during June 23rd–30th 2018 was dominated by the Saddleworth Moor and Winter Hill fires.

The economic cost of mortality caused by exposure to PM_2.5 from the fires is calculated using the 'Value of Prevented Fatality' (VPF) from the Department for Transport. (see section 2.4, table S4 and supplementary material section 6 for more details). Our results indicate the fires were associated with a £21.1 m economic cost between 23rd–30th June (95% CI: £10.7 m–31.2 m based on calculated excess mortality uncertainty (95% CI)) (see table 2(a) for more details). The estimates are broken down into human costs, medical costs, lost output and other costs (table 2(b)) (see section 2.4 for more details). This indicates the economic cost of the fires is dominated by the human cost (£13.9 m) and lost output (£7.0 m). The estimated economic cost of the fires suggests there are large economic gains to be made through the introduction of policies and education programmes to reduce the population's exposure to harmful air pollutants from fires.