The Northern California Camp Fire that took place in November 2018 was one
of the most damaging environmental events in California history. Here, we
analyze ground-based station observations of airborne particulate matter
that has a diameter
Wildfires have become increasingly prevalent in California. It has been
reported that, between 2007 and 2016, as many as 3672 fires occurred in
California, consuming up to 1759
Numerous studies have addressed wildfire events using a variety of model
frameworks and data sources (Shi et al., 2019; Herron-Thorpe et al., 2014;
Archer-Nicholls et al., 2015; Sessions et al., 2011). Shi et al. (2019) used
the Weather Research and Forecasting model online coupled with chemistry
(WRF-Chem) with Moderate Resolution Imaging Spectroradiometer
(MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS) fire data to study the wildfire of December 2017 in
Southern California. Herron-Thorpe et al. (2014) evaluated simulations of
the 2007 and 2008 wildfires in the Pacific Northwest using the Community
Multi-scale Air Quality (CMAQ) model with fire emissions generated by the
BlueSky framework and fire locations determined by the Satellite Mapping
Automated Reanalysis Tool for Fire Incident Reconciliation (SMART-FIRE).
That study suggested that underprediction of PM
The present study is a comprehensive investigation of air quality impacts of the Camp Fire using a combined analysis of ground-based and space-borne observations and WRF-Chem simulations. Descriptions of the observation and model are presented in Sect. 2; model evaluation is presented in Sect. 3; results of analysis are given in Sect. 4, followed by discussion and conclusion in Sect. 5.
The present study employs WRF-Chem (version 3.8.1) driven by the latest version of meteorological reanalysis data for initialization and boundary conditions. Fire emissions are determined by pairing active fire location data from the VIIRS satellite with the Brazilian Biomass Burning Emission Model (3BEM), which calculates species mass emissions from the burned biomass carbon density, combustion factors, emission factors, and the burning area. WRF-Chem simulations are evaluated against EPA surface observations and TROPOMI satellite products.
Study domain
The WRF-Chem simulation time period is 7 November 2018 (a day before the
fire began) to 22 November 2018 (when the fire was 90 % contained). We
carried out simulations over two domains (Fig. 1): domain d01 includes all of
California at
We use physical options of the Noah Land-Surface Model (Tewari et al., 2004), the Mellor–Yamada–Janjic (MYJ) boundary layer scheme (Janjic, 1994), and the Rapid Radiative Transfer Model (RRTM) (longwave) and Dudhia (shortwave) radiative transfer schemes (Dudhia, 1989). Cumulus parameterization is not included. The second-generation Regional Acid Deposition Model (RADM2) chemical mechanism coupled with the Modal Aerosol Dynamics model for Europe (MADE) and Secondary Organic Aerosol Model (SORGAM) (Zhao et al., 2011) are employed. Aerosol optical properties are calculated based on the volume approximation, for which the volume average of each aerosol species is used to calculate refractive indices (Jin et al., 2015). Aerosol radiative feedbacks on meteorology and chemistry are included in the simulations.
We use the National Emission Inventory for anthropogenic emissions (US EPA, 2018). Biogenic emissions are calculated online using the Guenther scheme (Guenther et al., 2006). Dust emissions are calculated online using the Goddard Chemistry Aerosol Radiation and Transport (GOCART) dust emission scheme with University of Cologne (UOC) modifications (Shao et al., 2011). Sea salt emissions are excluded. Technical details of wildfire emissions and the plume rise calculation are discussed in the next section.
Wildfire emissions are generated using the PREP-CHEM-SRC v1.5 preprocessor
(Freitas et al., 2011) employing 3BEM (Longo et al., 2010) with satellite data on detected fires. For
each pixel with fire detected, the mass of emitted species is calculated by
Plume rise model schematic. For each grid cell in which wildfire occurs, the plume rise model uses satellite fire products and the surrounding WRF-Chem environmental conditions to calculate two plume-top heights by using the land-type-dependent minimum and maximum wildfire heat fluxes. Smoldering-phase emissions are allotted to the surface layer, while flaming-phase emissions are distributed linearly aloft within the injection layers at a vertical resolution of 100 m.
The emission preprocessor generates a file formatted for WRF-Chem
containing the smoldering-phase surface emission fluxes of each species, the
fire size for each vegetation type, and flaming factor. Flaming factor is
the ratio of biomass consumed in the flaming phase to biomass consumed in
the smoldering phase. The 17 IGBP land cover classes are aggregated into
four main types: tropical forest, extratropical forest, savanna, and
grassland. The size of the wildfire and phase of combustion play important
roles in the structure of the plume and the vertical distribution of
emissions. Wildfire combustion is generally considered to occur in two
phases: smoldering and flaming. Emissions from the smoldering phase are
allotted to the first layer of the computational grid, while those from the
flaming phase are released at injection heights above the
surface, as determined by the plume rise model described below. Fire size
determines the total surface heat flux, as well as the entrainment radius of
the plume. Fire parameters are ascribed a daily temporal resolution and are
distributed to the WRF-Chem domains. The fire parameters are then input to
the plume rise model (Freitas et al., 2007, 2010). The plume rise model is a
one-dimensional model implemented in each WRF-Chem grid cell with an
independent vertical grid resolution of 100 m. It calculates the maximum
height to which a plume reaches and distributes emissions therein (Fig. 2).
The plume-top height, determined by the surface heat flux from the fire and
the thermodynamic stability of the atmospheric environment, is defined as
the height at which the in-plume parcel vertical velocity
Wildfire area by vegetation type in
Figure 3 shows the fire size and particulate matter emissions produced from
MODIS and VIIRS data. The Camp Fire burned primarily extratropical forest
vegetation (which comprised 68 % of the total burned area), followed by
savanna (23 % of total area). The flaming emission rate for species
The Fire Inventory from NCAR (FINN) version 1.5 (Wiedinmyer, 2011) is another fire emissions product that we will test in a sensitivity analysis. It is assembled for atmospheric chemistry models with a daily temporal resolution and a 1 km horizontal resolution. FINN is generated using satellite observations of active fires and land cover paired with emission factors and fuel loading estimates. The emissions are allocated to a diurnal cycle following WRAP (2005). FINN outputs the total wildfire emission flux, fire size, and land type fraction. As FINN does not include a smoldering- to flaming-phase ratio, the plume rise model calculates a ratio based on CO emissions.
The observational data include both ground-based measurements and satellite
observations. Meteorological and surface concentration data were obtained
from the NOAA's National Climatic Data Center (NCDC) and EPA Air Quality
System (AQS), respectively. We focus on three areas: the region closest to
the fire, the Sacramento Metro Area (population of 2.5 million), and the San
Francisco Bay Area (population of 7 million). Hourly observations of wind
speed at 10 m, wind direction at 10 m, temperature at 2 m, PM
To investigate the effects of key model parameters on the ability to predict the atmospheric impact of the wildfire, we conduct a range of sensitivity simulations. As meteorology and atmospheric structure play important roles in plume dynamics and the transport of particulate matter, we separately perturb the aerosol radiative feedback to meteorology, the planetary boundary layer parameterization, and the plume entrainment coefficient. To understand further the extent to which fire characteristics provided by satellite data can affect the simulations, we analyze the influence of fire data sources, the emission rate, and partitioning between smoldering-phase and flaming-phase emissions. A summary of these simulations is provided in Table 1.
Summary of sensitivity simulation setup.
Our evaluation focuses on the control simulation (S_CTRL). S_CTRL applies a factor of 3 to the smoldering emissions on 13 November and a factor of 2 to the smoldering emissions on 14–16 November due to the intermittent cloudy conditions over the Northern California on those days. S_CTRL uses the native flaming factor and fire size products, the default entrainment constant of 0.05, and the MYJ planetary boundary layer scheme. In the following scenarios, one parameter is individually perturbed from this configuration. S_EMRAW uses the native emissions input with unaltered smoldering-phase emissions, S_NOAERO turns off the aerosol radiative feedback to meteorological fields, S_FCTX2 doubles the flaming factor for the entire simulation period (thus increasing flaming-phase emissions without changing the smoldering phase), S_ENTR reduces the entrainment coefficient within the plume rise model from 0.05 to 0.02, and S_LSM employs an alternative land surface model and planetary boundary layer scheme. We perform another sensitivity simulation using FINN in place of VIIRS (S_FINN).
Comparison of AQS and NCDC wind observations (black) with
S_CTRL predictions (red) averaged over the three areas of
study:
Comparison of AQS and NCDC temperature observations versus
S_CTRL predictions:
The three spatial areas of our interest differ significantly in topography
and meteorology. Figure 4 shows the averaged wind observations and
S_CTRL predictions. S_CTRL captures general
wind patterns and achieves strong correlation with observed temperatures in
each of the areas (Fig. 5). In the first few days of the Camp Fire, the
foothills and the Sacramento area experienced strong northerly winds, while
the Bay Area experienced northeasterly winds, both predicted by the
simulation. Other distinct features like those on 11 November near the fire
and in the Bay Area are also reproduced by S_CTRL with some
bias in timing. In the Bay Area, winds were typically southerly at speeds
less than 2
Summary of meteorological model performance metrics for the simulation duration.
Comparison of AQS surface PM
Figure 6 shows the predicted evolution of surface PM
Summary of model performance metrics for surface PM
Comparison of AQS surface black carbon (
Comparison of TROPOMI UV aerosol index and S_CTRL total BC column during 8–18 November at 13:30 LT as a proxy for plume structure and motion. Due to cloud coverage, no data for 15 November are shown. Positive aerosol index (warm colors) indicates aerosols that absorb radiation like black and brown carbon. The spatial distribution of the plume is generally captured on most days. The simulation also captures some of the finer structures seen by the satellite, though they are somewhat displaced.
Error in surface PM
To study the structural evolution of the wildfire plume, we compare
simulated total black carbon column with TROPOMI UVAI satellite retrievals
(Fig. 8). TROPOMI UVAI is based on the difference between
wavelength-dependent Rayleigh scattering observed in an atmosphere with
aerosols and that of a modeled molecular atmosphere (Stein Zweers et al.,
2018). This difference is measured in the UV spectral range where ozone
absorption is small. A positive residual (red coloring) indicates the
presence of UV-absorbing aerosols, like black carbon (BC), while a negative
residual (blue coloring) indicates presence of non-absorbing aerosols. As
WRF-Chem does not generate an aerosol index parameter, we compare UVAI to
total BC column, a significantly absorbing aerosol. Over the period of the
simulation, broad characteristics and shape, as well as some more distinct
features, of the Camp Fire plume are reproduced by S_CTRL.
Using similar input data sources and WRF-Chem configuration but a simpler
plume rise model, Shi et al. (2019) also capture the general shape of the
plume but underestimate aerosol magnitude. Discrepancies in
S_CTRL plume transport correlate to bias in surface
PM
Vertical profile of PM
Surface PM
To investigate the predicted decrease of surface PM
The TROPOMI aerosol layer height (ALH) retrieval represents vertically localized aerosol layers
within the free troposphere in cloud-free conditions and is designed to
capture aerosol layers produced by biomass burning aerosol (such as
wildfires), volcanic ash, and desert dust (Apituley et al., 2018). ALH is
retrieved based on the significant effect of aerosol vertical structure on
the high-spectral-resolution observations in the
Comparison of TROPOMI aerosol layer height
We compare the satellite-derived aerosol layer height to WRF-Chem
predictions of PM
We conduct sensitivity simulations to investigate the effects of various parameters on the ability of the WRF-Chem model to accurately predict downwind PM concentrations from wildfires. As meteorological conditions and related boundary structure play important roles in plume dynamics and the transport of PM, we separately test the aerosol feedback to meteorology and the land surface model. To understand the extent to which fire characteristics provided by satellite data can affect the simulation, we analyze the fire product sources (VIIRS versus FINN), the total fire emissions, and the division between smoldering-phase and flaming-phase emissions. To examine the influence of the plume rise model, we perturb a key parameter: the entrainment coefficient.
Comparison of meteorology generated by S_CTRL
(solid red line) and S_NOAERO (in which aerosol effects do not
feed back to the meteorology; dashed blue line) over the three areas of study:
By absorbing and scattering solar radiation, aerosols can impact the
radiative fluxes, cloud formation, and precipitation in the atmosphere (Wang
et al., 2016, 2020), and, in turn, the meteorological conditions for aerosol
formation, transport, and removal (Li et al., 2019). WRF-Chem has the option
to couple aerosol–radiative direct effects with meteorology simulation.
S_NOAERO uses the same input data and configuration as
S_CTRL but disables the aerosol radiative feedback. Figure 12 shows
the evolution of surface wind speed and temperature throughout the
wildfire near the source (Fig. 12a), in Sacramento (Fig. 12b), and in the Bay Area (Fig. 12c).
The aerosol radiative impact on simulated meteorology is more pronounced for
surface temperature than wind. When aerosol radiative feedbacks are
noticeable, colder temperatures and calmer winds are found near the surface.
Generally, feedbacks are more evident in the region closer to the fire
sources with larger PM concentrations. Also, in the Bay Area, the largest
changes in meteorology coincide with the largest differences in surface
PM
Time series of surface PM
Currently, fire emission inventories generally have large uncertainty. Although wildfires have been studied for decades and there is vast literature characterizing biomass combustion emissions, there are large knowledge gaps in the composition of these emissions when a nontrivial fraction of the burnt area includes built environment comprising a vast array of non-biomass-related materials. For the Camp Fire, there is a paucity of the types of burned land cover and fire emissions data required to incorporate these considerations into model simulations. WRF-Chem input fire files produced with VIIRS and PREP-CHEM-SRC include fire size, smoldering emission flux, and flaming factor. Here, we test the sensitivity of predictions to different emission dataset (FINN (S_FINN) versus VIIRS/MODIS), as well as emission injection parameters, such as the smoldering emission flux (S_EMRAW) and flaming factor (S_FCTX2). S_FINN produces very little aerosol, though it captures the timing of some peaks. The aerosol underestimation may be a result of bias in the emission inventory or an issue of its implementation in the plume rise model code, as FINN specifies total wildfire emissions rather than a smoldering and flaming distribution.
When the VIIRS emission inventory is used, the total wildfire emission flux can
be altered through two parameters: the smoldering emission flux at the
surface and the flaming factor. Directly increasing the smoldering emission
flux adds emissions to the surface layer and increases flaming-phase
emissions proportionally. Figure 13 shows the impact of doubling smoldering
emissions on 13 November and tripling them during 14–16 November. These
changes to the inventory more than double concentrations of surface
PM
The plume rise model parameterizes entrainment as proportional to the plume
vertical velocity and inversely proportional to the plume radius (Freitas et
al., 2010). Greater entrainment causes rapid cooling, such that near-surface
plume temperatures are only slightly warmer than the environment, lowering
buoyancy and reducing the plume height. Larger wildfires generate less
entrainment and reach higher injection heights. The parameterization also
includes the effect of horizontal winds on entrainment. Strong wind shear
can enhance entrainment and increase boundary layer mixing (Freitas et al.,
2010). Archer-Nicholls et al. (2015) decreased the original entrainment
coefficient (Freitas et al., 2007) from 0.1 to 0.05 to improve their
simulations of a wildfire. As the Camp Fire developed rapidly and intensely,
we performed the sensitivity simulation S_ENTR with a lower
entrainment coefficient of 0.02 to allow for higher injection heights.
However, entrainment perturbation resulted in less than 1 % change in
surface PM
Comparison of surface PM
We compare simulations using two different land surface models (LSMs) which include the planetary boundary layer (PBL) schemes: the Noah LSM with MYJ PBL and the Pleim–Xiu LSM (referred to here as P-X) with the Asymmetric Convection Model 2 (ACM2) PBL (Janjic, 1994; Pleim and Xiu, 1995; Chen and Dudhia, 2001; Pleim, 2007). Land surface models simulate the heat and radiative fluxes between the ground and the atmosphere (Campbell et al., 2018). The Noah LSM has four soil moisture and temperature layers, while the P-X LSM has two (Hu et al., 2014; Campbell et al., 2018). Both include a vegetation canopy model and vegetative evapotranspiration. The PBL scheme provides the boundary layer fluxes (heat, moisture, and momentum) and the vertical diffusion within the column. It uses boundary layer eddy fluxes to distribute surface fluxes and grows the PBL by entrainment. A key feature of PBL schemes is the inclusion of local mixing (between adjacent layers) and/or non-local mixing (from the surface layer to higher layers). The MYJ scheme is a turbulent kinetic energy prediction, while the ACM2 scheme is a member of the diagnostic non-local class. MYJ solves for the total kinetic energy in each column from buoyancy and shear production, dissipation, and vertical mixing. ACM2 has two main components: a term for local transport by small eddies and a term for non-local transport by large eddies. Coniglio et al. (2013) showed that the MYJ scheme can undermix the PBL in locations upstream of convection in the presence of overly cool and moist conditions near the ground in the daytime, whereas ACM2 can result in an excessively deep PBL in evening. Pleim (AMS, 2007) also noted that ACM2 predicts the PBL profile of potential temperature and velocity with greater accuracy.
The use of P-X and ACM2 results in substantially different aerosol trends
and plume evolution, the effects of which are largely location-dependent
(Fig. 13). Near the fire and in the Bay Area, S_LSM produces
little similarity in surface PM
To test the linearity of different factors in regulating the fire-related PM
pollution, we choose two factors, emission flaming factor and aerosol
radiative feedback, and conduct a new experiment by jointly perturbing these
two. We compare the results from this joint perturbing experiment with those
from each individual perturbing experiment and the linear sum of the two in
Fig. 14. It shows that for the most times, the effect of joint
perturbation is close to the sum of the two individual effects (the black
line follows well with the black circles), indicating that the relatively
good linearity and additivity hold between those two factors in a general
sense. The exception occurs under the extreme conditions. During 14–18 November
when the plume was thick and PM
The record-breaking Camp Fire ravaged Northern California for nearly 2 weeks.
At a distance of 240 km downwind of the wildfire, Bay Area surface
PM
We focus on three geographic areas: the vicinity of the wildfire,
Sacramento, and the San Francisco Bay Area. The control experiment was able
to simulate the general transport and extent of the plume as well as the
magnitude and temporal evolution of surface PM
Future studies are needed to further improve the present modeling framework
to simulate wildfires. Some wildfires exhibit a distinct diurnal cycle, but
the current fire preparation module has not utilized the time information of
the fire radiative power measurements by the polar-orbiting satellites.
Also, the current land cover and vegetation type data are still relatively
coarse in spatial resolution and classification accuracy, which cannot fully
resolve a small town in a rural area. In fact, the Camp Fire reportedly
burned the town of Paradise, California, between 8 and 10 November 2018. The
town of Paradise covered 47
The recent TROPOMI aerosol layer height product shows promise as an analytical tool but requires further development of the method by which it can be directly compared to WRF-Chem. Given the assumptions required to perform the TROPOMI ALH retrieval, more research is needed to compare that product with any height retrievals from MODIS/MAIAC (Lyapustin et al. 2020), MISR, and CALIPSO. The intercomparison can help quantify measurement uncertainty. Herron-Thorpe et al. (2014) noted that careful consideration must also be given to the vertical coordinates across models and satellite products, as discrepancies in reporting heights in reference to sea level, ground level, or the geoid can influence analyses.
WRF-Chem model code is available for download via the WRF website
(
US Environmental Protection Agency Air Quality System Data Mart (internet
database) is available for download (
YW, JHS, and JHJ conceived and designed the research. YW and BR performed the WRF-Chem simulations. BR, YW, and JHS performed the data analyses and produced the figures. BZ provided technical support for fire emission preparation. ZCZ helped satellite data analyses. BR, YW, and JHS wrote the paper. All authors contributed to the scientific discussions and preparation of the manuscript.
The authors declare that they have no conflict of interest.
This study was supported by the Jet Propulsion Laboratory, California Institute of Technology, under contract with NASA. We thank Kristal R. Verhulst, Yi Yin, Don Longo, Gonzalo Ferrada, and Saulo Freitas for their support and discussion.
This study has been supported by the AQ-SRTD project at the Jet Propulsion Laboratory, California Institute of Technology, under contract with NASA, and the NASA ACMAP, CCST, and TASNPP programs.
This paper was edited by Joshua Fu and reviewed by two anonymous referees.