3.1 Quality filtered XCO and comparison to CAMS

As a result of the Camp and Woolsey fires ignited on 8 November, associated smoke overcast large parts of the state for nearly 2 weeks. This can clearly be seen on the VIIRS true colour images in Fig. 1 for the first days of the fires. Sentinel-5 Precursor and Suomi-NPP fly in loose formation, with Sentinel-5P trailing behind by 3.5 min, ensuring that both satellites observe (almost) the same scene. Thus, the corresponding images can be compared directly.

Figure 2Performance of the implemented quality filter for the example day 31 October. (a) Cloud-free reference XCH₄ abundances (5–7 November). (b) Unfiltered XCH₄ data. (c) XCO after application of the filter removing scenes with unrealistic low XCH₄. (d) Comparison of the standard quality filter (SQF, 1: excluded) with the VIIRS cloud classification (1: cloudy). Matching classifications are shown in white and green (agreement with VIIRS: 78 %, passing SQF: 32 %). (e) As before but for the alternative quality filter (AQF) used in the presented analysis (agreement with VIIRS: 81 %, passing AQF: 39 %). (f) VIIRS true colour image.

The performance of the quality filter based on simultaneously measured methane is demonstrated in Fig. 2 for an example day before the start of the analysed fires. In line with the error analysis based on synthetic data presented in Schneising et al. (2019), there is typically a considerable underestimation of XCH₄ in the presence of clouds due to shielding of the underlying partial columns. After application of the alternative quality filter, there are no obvious issues with the XCO data. The relaxed quality screening algorithm provides similar results as the standard filter concerning the overall agreement rate with the VIIRS cloud product but actually yields more scenes passing the filter (about 20 % for the analysed example day).

Figure 3Retrieved methane column-averaged mole fractions from TROPOMI for the same days as in Fig. 1. XCH₄ is used to filter out scenes with significant underestimation (dotted scenes), mainly due to reduced near-surface sensitivity in the presence of clouds or smoke due to shielding of the subjacent partial columns. The Central Valley exhibits combined anthropogenic methane emissions from oil fields and agriculture (Schneising et al., 2019).

Figure 3 shows the daily XCH₄ distribution over California, which is used for quality screening. For each day, XCH₄ resembles the prefire background abundances shown in Fig. 2a with the exception of considerable underestimations here and there mainly due to reduced near-surface sensitivity in the presence of clouds or smoke near the origin of the fires. However, at a sufficient distance from the seat of the fire the XCH₄ abundances are not affected and a quantitative analysis is still possible, even in cases in which efficient scattering in the visible spectral range is indicated by extensive plumes in the VIIRS images. The explanation for this is the particle size distribution of the wildfire smoke. While clouds typically consist of water droplets with an effective radius of the order of 10 µm, smoke is dominated by considerably smaller particles. The mass distribution of smoke plumes shows a prominent peak at about 0.3 µm (Stith et al., 1981) but is nevertheless dominated by a small number of supermicron-sized particles, including some very large particles (Radke et al., 1990). As a consequence of the different size distributions, clouds have a typical Ångström exponent α=0 and thus no wavelength dependence of the aerosol optical depth, while biomass burning aerosols have a distinct wavelength dependence with typical α ranging between 1 and 2 depending on the fire (Eck et al., 2009). The submicron particles reduce visibility and lead to extended smoke plumes over large distances in the VIIRS true colour reflectances shown in Fig. 1. However, the 2.3 µm spectral range, in which the satellite measurements are taken, is subject to little scattering at these small particles. The satellite retrievals close to the source of the fire are rather affected by the large supermicron-sized particles, which have a short atmospheric lifetime, tend to fall out rapidly (World Health Organization, 2006), and thus become more and more negligible when departing from the seat of the fire. Thus, a reliable XCO retrieval is possible in smoke plumes in the far field of the fire origin for scenes passing the quality filter. Corresponding simulations with a realistic aerosol optical depth and Ångström exponent are included in the error analysis in the next subsection to quantify the impact of scattering at smoke aerosols.

Figure 4Retrieved carbon monoxide column-averaged mole fractions from TROPOMI for the same days as in Fig. 1. Dotted scenes are excluded by the quality filter based on simultaneously retrieved XCH₄. Also shown is the mean wind in the boundary layer obtained from ECMWF data.

Figure 4 shows the XCO distribution over California, which matches the smoke emission and transport patterns detected by VIIRS unambiguously. This substantiates the fact that the observed CO enhancements are actually originating from the wildfires. It can be seen that the abundances over the major cities we want to analyse are typically not filtered out and are thus suitable for a quantitative analysis. However, the quantitative interpretation of scenes right above or too close to the origin of the fire is limited by reduced vertical retrieval sensitivity near the surface, and they are consequently filtered out.

Figure 5CAMS near-real-time CO analysis for the first days of the fires at 24:00 UTC, corresponding to 16:00 local time (Pacific Standard Time).

For comparison, Fig. 5 shows the CAMS near-real-time CO analysis on a $\mathrm{0.1}{}^{\circ }\times \mathrm{0.1}{}^{\circ }$ grid for the same days shown in the previous figures and the closest available time to the TROPOMI overpass at 13:30 local time. As CAMS is available in time steps of 6 h, the analysis corresponding to 16:00 local time is used for the comparison. Although CO emissions from the fires are obviously included in the CAMS data, the transport patterns seem to be somewhat different. While the patterns are broadly consistent for 9 and 10 November, the modelled wind fields close to the fires seem to deviate on 8 and 11 November, which results in a longer continuance of the plume over land, while the VIIRS images and the TROPOMI data suggest a faster transport westwards to the sea. This can also be seen in Fig. 6, showing departures of the CAMS analysis from the TROPOMI XCO after averaging the satellite data on the CAMS resolution. Apart from the partially different transport patterns, the intensity distribution close to the fire sources is also different, with CAMS abundances being considerably higher for the most part and deviations reaching several hundred parts per billion (ppb). This may be due to overestimated wildfire fluxes, underestimated initial horizontal transport in the vicinity of the fire sources, or a combination of both.

Figure 6Difference of CAMS XCO analysis to TROPOMI/WFMD satellite measurements.

3.2 Boundary layer concentration anomalies and associated uncertainties

To assess the CO burden in Californian cities, the boundary layer CO concentration anomaly is computed according to Eq. (1). The diurnal variation of the ECMWF ERA5 boundary layer heights and their inherent uncertainties are illustrated in Fig. 7. There is a strong diurnal cycle, with low values at night and maximal values around local noon close to the time of the TROPOMI overpass at 13:30. The boundary layer concentration uncertainty arising from boundary layer height σ(h_bl) is determined from the maximal ERA5 ensemble uncertainty between 13:00 and 14:00 local time and the variation within this hour in each case. Typical values of σ(h_bl) range between 10 % and 25 %. The ERA5 boundary layer height around the satellite overpass time is about 300 m smaller than the boundary layer height derived from aerosol and turbulence detection lidar measurements by different retrieval methods (Wang et al., 2019). This uncertainty associated with the estimation method is additionally taken into account in the error budget by an extra term σ_m(h_bl) quantifying the percentage impact on the smaller end of Δρ_bl.

Figure 7Diurnal variations of the boundary layer heights obtained from the ECMWF ERA5 reanalysis for major Californian cities (solid) and fires (dotted). The uncertainty estimates are based on a 10-member 4D-Var ensemble; the additional uncertainty associated with the estimation method of the boundary layer height is not shown here. Also shown are mean IS4FIRES smoke injection heights and their variation for both analysed fires as horizontal bars and hatched areas. The grey-shaded area illustrates the TROPOMI overpass time. On 8 November the injection height of the Woolsey Fire is zero because it started later in the day.

Download

The potentially largest source of uncertainty in the calculation of the boundary layer CO burden is plume dynamics and the question of whether all CO remains in the boundary layer or if a certain proportion reaches the free troposphere. The vertical distribution of emissions near the source is driven by the fire radiative power and the local ambient atmospheric conditions such as stability and humidity. Three types of wildfire plumes are distinguished by the amount of condensed water vapour during plume formation (Fromm et al., 2010): (1) dry smoke plumes, which contain water vapour and usually stay within the boundary layer, (2) pyrocumulus containing water droplets either staying in the boundary layer or reaching the free troposphere depending on atmospheric conditions, and extreme (3) pyrocumulonimbus scenarios containing ice particles and potentially reaching the stratosphere. Typically, most of the biomass burning emissions stay within the mixing layer, and cases with pyro-convection or direct injection to the free troposphere or even higher are rare (Labonne et al., 2007; Mazzoni et al., 2007; Tosca et al., 2011).

As can be seen in Fig. 7, the IS4FIRES injection heights corresponding to the top of the plume are equal to or smaller than the respective maximum boundary layer height at the location of the fires (all the more when considering the additional uncertainty associated with the estimation method of the boundary layer height, which is not shown in the figure), with the exception of the first day of the Camp Fire. This sole discrepancy may be linked to overestimated fire radiative power for the Camp Fire on the day of ignition, which is also suggested by the comparison of the CAMS XCO analysis to the TROPOMI retrievals, showing considerably higher abundances for CAMS in the vicinity of the fire source. In summary, the IS4FIRES injection height analysis indicates that most of the CO load stays within the boundary layer. Furthermore, the entire state of California was at least abnormally dry within the analysed time period, with a moderate drought at the Camp Fire origin and a severe drought at the seat of the Woolsey Fire according to the United States Drought Monitor (https://droughtmonitor.unl.edu/, last access: 20 January 2020). These are favourable conditions for dry smoke plumes being trapped in the boundary layer, also rendering later deep moist convection with transport to the free troposphere during the first days of the fire at another location unlikely. Finally, there is also no indication for pyrocumulus or pyrocumulonimbus in the VIIRS true colour images as there is no obvious cloud formation over the fires (see Fig. 1).

Nevertheless, partial venting to the free troposphere cannot be entirely excluded and we therefore introduce an uncertainty σ(h_inj) arising from unknown plume dynamics, which is only applied to the smaller end of the boundary layer concentration anomaly because lost CO of this type can only lead to an overestimation of the near-surface concentrations. σ(h_inj) for the analysed cities is estimated by the CO mass fraction f_a above the upper bound of the ERA5 boundary layer height uncertainty range as determined from the CAMS model CO vertical profiles (see Fig. 8). As there is no indication that the CO fraction in the free troposphere grows significantly during the analysed period over the cities considered, mean CAMS profiles for days with substantial CO enhancement are examined. The CAMS CO profile analysis further reinforces the assumption that most of the emitted CO stays within the boundary layer even 4 d after ignition and at a greater distance from the fire sources. f_a is below 10 % with the exception of San Diego. Thus, the uncertainty σ(h_inj) is set to 30 % for San Diego and 10 % for the other cities.

Figure 8Mean CAMS CO profile enhancement relative to the prefire background (7 November). Fresno is not shown because there is no significant enhancement during 8–11 November in the CAMS model. The upper edges of the grey areas represent the surface pressure, the black vertical lines illustrate the upper bound of the uncertainty range of the estimated mean ERA5 boundary layer height, and f_a is the CO mass fraction above the boundary layer.

Download

Another potential error source associated with fires is smoke aerosol. Scenes with reduced near-surface sensitivity due to clouds and smoke with large particles near the seat of the fires are automatically filtered out using simultaneously measured methane. Figure 3 also demonstrates that methane is not considerably increased compared to the prefire background abundances (Fig. 2a) and that the XCO enhancement patterns are not resembled in XCH₄. Thus, it can be excluded that the detected XCO enhancement is only an artefact as a result of light path lengthening because of aerosol scattering at the particulate matter of the smoke because such systematic errors would affect both retrieved gases similarly.

To assess the potential impact of smoke aerosol quantitatively, simulated measurements are used. This means that sun-normalised radiances for an assumed smoke scenario are calculated with the radiative transfer model, which are subsequently used as measurement input in the retrieval. The errors are then defined as the deviation of the retrieved columns for the smoke scenario from the corresponding columns for the background scenario also used to calculate the forward model lookup table. To model wildfire conditions at a sufficient distance from the seat of the fire with low visibility but decreasing scattering issues at larger wavelengths (consistent with Figs. 1 and 3) we use the extreme in BL aerosol scenario originally introduced in Schneising et al. (2008), containing urban aerosol with a significant soot fraction (Shettle and Fenn, 1979), combined with an extreme CO profile with a 10-fold enhancement in the boundary layer compared to the standard profile. The aerosol scenario used (aerosol optical depth τ_550 nm≈3 and Ångström exponent α≈1) is considered a realistic worst-case scenario for the analysed fires because it is at the upper end of optical depths and at the lower end of Ångström exponents for typical fire aerosols (Eck et al., 2009). Furthermore, the corresponding aerosol profile is consistent with the previous results for the vertical distribution of the emitted species during the first 4 d of the fires. Thus, the Camp Fire and the Woolsey Fire very likely exhibit less scattering in the 2.3 µm spectral range than our model scenario assumes, at least during the period analysed. The corresponding results are summarised in Fig. 9. Typical systematic CO errors for Californian cities on the analysed days range between about −3 % and 2 % for the assumed aerosol type and CO profile. Therefore, the uncertainty due to smoke aerosol σ(a_smo) is set to 5 %, adding an extra amount due to the uncertainty of the actual aerosol type.

Figure 9Panel (a) shows the aerosol extinction profiles used in the analysis of smoke aerosol errors. Also given are the corresponding aerosol optical depths at 550 nm. The other panels show the systematic errors of CO and CH₄ as a function of solar zenith angle and albedo when using the extreme instead of the background aerosol scenario. The green boxes highlight the typical conditions for the Californian cities on the analysed days using percentiles corresponding to 1σ (68 % of data) and 2σ (95 % of data). The green circle is the pair of median albedo and median solar zenith angle.

Download

The total uncertainty of the boundary layer concentration anomaly σ(Δρ_bl) is determined by

Averaged boundary layer concentration anomalies of CO (relative to 7 November) in major Californian cities during the first days of the Camp and Woolsey fires are presented in Fig. 10 together with the total uncertainty of Eq. (2) and an estimate of the areal variation measuring the inhomogeneity of the CO concentrations within the city area. The largest values are found for Sacramento and San Francisco on 9 and 10 November due to the prevailing wind conditions, with boundary layer concentration anomalies of about 2.5 mg CO m⁻³, which is well below the national CO air quality standard of 10 mg m⁻³ even after adding a typical background concentration of about 0.5–1.0 mg CO m⁻³. The cities in the southern part of the state are less affected owing to more favourable weather conditions.

Figure 10Averaged boundary layer concentration anomalies of CO (relative to 7 November) and associated areal variations and uncertainties (1σ) in major Californian cities during the first days of the Camp and Woolsey fires.

Download

Although the Sacramento and San Francisco city averages are compliant with air quality standards, the large associated areal variations indicate an uneven CO distribution within both towns, in particular for Sacramento. This interpretation is supported by the CO distribution depicted in Fig. 4, showing that the plume's edge of the Camp Fire is located near Sacramento, leading to a larger burden in the northwest compared to the rest of the city.

Figure 11Boundary layer concentration anomalies of Sacramento and its environs determined from TROPOMI CO total column measurements and boundary layer heights from the ECMWF ERA5 reanalysis. Highlighted are the satellite scene with the maximal city area value (5.42 mg CO m⁻³ [3.41–6.00; 1σ], red) and the location of the AQMIS sites in North Highlands and at Del Paso Manor (black). The anomalies based on the maximum values of the ground-based sites (3.60 and 3.00 mg CO m⁻³) are colour-coded in the inner circle at the site location; the anomalies based on the daily averages (1.93 mg CO m⁻³ each) are colour-coded in the outer circle.

The largest burden with respect to CO within all city radii is actually found on 10 November about 10 km to the east of Sacramento International Airport, where one finds a considerable column enhancement of 3.14 g m⁻². Given the ECMWF ERA5 boundary layer height of 580 m, this corresponds to a boundary layer concentration anomaly of 5.42 mg CO m⁻³ [3.41–6.00; 1σ]. The largest enhancement on 9 November is also located in the vicinity of Sacramento Airport (about 10 km to the southwest) and amounts to 3.13 g m⁻², with a boundary layer height of 592 m leading to a boundary layer concentration anomaly of 5.28 mg CO m⁻³ [3.32–5.93; 1σ]. Thus, the national ambient air quality standard of 10 mg CO m⁻³ was likely not exceeded even for the most polluted city scenes and after adding a typical background of about 0.5–1.0 mg CO m⁻³.

To further assess the described area with significantly increased boundary layer concentrations, we revisit the discussed contaminated scene near Sacramento International Airport on 10 November and analyse associated results from CAMS and ground-based air quality monitoring data provided by the California Air Resources Board. For the grid box comprising the mentioned satellite scene, CAMS predicts a considerably larger column enhancement of 5.93 g m⁻², corresponding to a boundary layer concentration anomaly of 10.23 mg CO m⁻³ [6.51–11.04; 1σ] using the ERA5 boundary layer height and the associated uncertainties. Although this is almost twice as high as the satellite-derived concentration anomaly and potentially exceeds the national ambient air quality standards, the error bars of the CAMS and satellite-derived concentration anomalies almost overlap due to the relatively large uncertainties arising from boundary layer height estimation and unknown plume dynamics.

Ground-based measurements are available from the Air Quality and Meteorological Information System (AQMIS) network (California Air Resources Board, 2018), providing daily maximum and daily average values. There are three CO measurement sites in Sacramento County. For the site at Bercut Drive in Sacramento the data set is incomplete during the first days of the fire and therefore excluded from the comparison. The second site at Blackfoot Way in North Highlands is located farther north and closer to the analysed contaminated satellite scene. The maximum value during the first 4 d of the fire is stated to be 4.1 ppm (4.8 mg CO m⁻³) on 10 November. Relative to the maximum value of 7 November this corresponds to a concentration anomaly of 3.60 mg CO m⁻³. The third site is at Del Paso Manor in Sacramento with a maximum value of 3.8 ppm (4.4 mg CO m⁻³) on 10 November corresponding to a concentration anomaly of 3.00 mg CO m⁻³. When using the daily averages instead of the maximum values, the concentration anomalies amount to 1.93 mg CO m⁻³ for both sites.

Figure 11 shows the boundary layer concentration anomalies in Sacramento and its surrounding districts, allowing for an overview of the situation by highlighting the locations of the different measurements. As can be seen, the AQMIS sites are located east of the satellite scene with the maximal city area CO value, where concentrations are beginning to decline steeply. The corresponding satellite averages at both analysed AQMIS sites are broadly consistent with the ground-based measurements, taking into account the potential variability within a satellite scene indicated by the scene-to-scene gradient of the satellite data and the fact that the sites are located at the edge of satellite scenes. While the ground-based anomaly based on the maximum values in North Highlands matches well with the value of the associated satellite scene, the ground-based anomaly based on the daily averages rather resembles the values of adjacent satellite scenes to the east or to the south. At Del Paso Manor the opposite is true: the ground-based anomaly based on the daily averages fits the surrounding satellite scene well, while the anomaly based on the maximum values rather matches the adjacent satellite scene to the north.