Spatially and temporally resolved measurements of NO x ﬂuxes by airborne eddy covariance over Greater London

Abstract. Flux measurements of nitrogen oxides (NOx) were made over London using
airborne eddy covariance from a low-flying aircraft. Seven low-altitude flights were conducted over Greater London, performing multiple overpasses across the city during eight days in July 2014. NOx fluxes across the
Greater London region (GLR) exhibited high heterogeneity and strong diurnal
variability, with central areas responsible for the highest emission rates
(20–30 mg m−2 h−1). Other high-emission areas included the M25 orbital motorway. The complexity of London's emission characteristics makes it challenging to pinpoint single emissions sources definitively using
airborne measurements. Multiple sources, including road transport and
residential, commercial and industrial combustion sources, are all likely to contribute to measured fluxes. Measured flux estimates were compared to
scaled National Atmospheric Emissions Inventory (NAEI) estimates, accounting
for monthly, daily and hourly variability. Significant differences were found between the flux-driven emissions and the NAEI estimates across
Greater London, with measured values up to 2 times higher in Central London than those predicted by the inventory. To overcome the limitations of
using the national inventory to contextualise measured fluxes, we used
physics-guided flux data fusion to train environmental response functions
(ERFs) between measured flux and environmental drivers (meteorological and surface). The aim was to generate time-of-day emission surfaces using
calculated ERF relationships for the entire GLR; 98 % spatial coverage was achieved across the GLR at 400 m2 spatial resolution. All flight leg
projections showed substantial heterogeneity across the domain, with high
emissions emanating from Central London and major road infrastructure. The
diurnal emission structure of the GLR was also investigated, through ERF,
with the morning rush hour distinguished from lower emissions during the early afternoon. Overall, the integration of airborne fluxes with an
ERF-driven strategy enabled the first independent generation of surface
NOx emissions, at high resolution using an eddy-covariance approach,
for an entire city region.



S1.1 Instrument information
. Instrument schematic for fast AQD dual-channel chemiluminescence NOx analyser (Fast-AQD-NOx). Dotted flow path represents zero count-rate flow path for both channels, giving a zero-count rate for each PMT. pressure is kept at >300 hPa to negate the effect of changing altitude on instrument sensitivity. The blue-light converter consists of a Teflon block containing a milled cavity (10 mL volume) down its centre, and three 395 ± 20 nm wavelength LED positioned on either side (Reed et al., 2016). Below 400 nm NO2 photolytically degrades to NO and ground-state molecular oxygen (O[ 3 P]). The converter is cooled using a Peltier cooler to reduces thermal interference products and exhibits a conversion efficiency of >85 % and a resonance time of 0.11 s.
The dark count rate on each PMT was assessed using the zero-volume flow path shown in Fig. 1s as the dotted line. Mixing between the sample flow and O3 occurs in the zero-volume, ensuring the chemiluminescence reaction occurs before the sample reaches the reaction vessel. Typical PMT dark counts ranged from 2,000-6,000 counts s -1 . The dark-count rate on each PMT was measured frequently and subtracted from the ambient signal to give an accurate background correction. Instrument design takes into account water vapour quenching with regards to photon counting (Ridley et al., 1992). A constant amount of deionised water vapour (d.H2O) is added to the O3 supply, increasing the concentration within the reaction vessel to ~32 parts per thousand (ppth). By standardising water vapour within the reaction vessel, any changes in atmospheric water vapour concentration are negated which negatives significant chemiluminescence quenching effects.
The sensitives of both the PMTs and the blue-light converter can drift over time, requiring calibration to a known NO/NO2 standard. A 5 ppm NO standard (BOC Group plc., supplied and certified) was used to calibrate against, which was further certified against a high accuracy National Physical Laboratory (NPL) standard, before and after the field campaign. Instrument mass-flow controllers were calibrated before and after field campaigns using a gilibrator (high accuracy electric bubble flow meter). PMT sensitivity was determined by standard addition of a small flow (5 -10 sccm) of NO calibrant gas to a flow of NOx-free air. NOx free air was obtained either by flying above the boundary layer where NOx levels are low or by removal using a Sofnofil/charcoal trap attached to the sample inlet. This gave a NO mixing ratio in the range 5 -10 ppt. NO channel sensitivity values range from 7 -8 counts ppt -1 . NO2 detector sensitivity was also determined by the same method, with typical value ranges from 9 -11 counts ppt -1 . In addition to detector sensitivity, the conversion efficiency of the blue-light converter was also assessed. A known NO2 mixing ratio was generated by titration reaction of NO calibrant gas with O3, generated by a 5 sccm flow of O2 passing by a 254 nm Oriel Instruments Mercury UV Pen-Ray lamp. The converter was found to give > 85% conversion efficiency during the entire campaign.

S1.2.1 eddy4R Software
The eddy4R software is a family of R packages (Core Team, 2019) which create a modular function-based software solution for EC data processing, as described in Metzger et al., (2017). A development and systems operation approach (DevOps) was utilised to create reproducible, open-source, and extensible software that is version controlled. This DevOps schema enables a release and iteration cycle that, to date, has yielded the eddy4R.base, eddy4R.qaqc, and eddy4R.stor packages on a publicly available GitHub repository (Metzger et al., 2019;Xu et al., 2019). This modular framework facilitates scientific communitydriven code development that extends the eddy4R software suite's capabilities. In the present study, we extended eddy4R to handle fluxes from a wide variety of chemical species by adding to the eddy4R.turb package that is currently in development.
To ensure portability and reproducibility, the eddy4R packages integrate into a Docker image, which builds upon a Linux computational environment and resolves all system and R dependencies (https://www.docker.com/why-docker). The Docker image hardens the code against operating system-induced anomalies and streamlines the code base to the essential requirements for processing. GitHub automatically triggers, and version controls Docker image builds, which are housed on Dockerhub.
Continuous integration testing through Travis-CI and subsequent code reviews complete the build chain. This fosters rapid code development across teams and functionalities while mitigating unintentional errors that could corrupt the main codebase.
The eddy4R-Docker DevOps framework thus provides a foundation for the distributed development of novel algorithmic solutions and their scalable execution. The described approach provides the end-user with a practical approach towards version control and result reproducibility.

S1.2.2 Wavelet Eddy Covariance
Wavelet EC uses continuous wavelet transformation (CWT) to extract time-frequency or space-wavenumber information from atmospheric signals. For this study, the Morlet wavelet (Cohen, 2019) was chosen due to its strong track record in quantifying atmospheric turbulence (Karl et al., 2013;Thomas and Foken, 2005). The complete covariance between two signals (x & y) is deduced by examining global covariance across all frequency scales, as shown in Eq. (1). aj defines the frequency domain scales, bn the time-domain scales, δt the steps between time-domain scales, δj the spacing between frequency domain scales, length of the data series (N) and Cδ wavelet specific reconstruction factor. CWT frequency scales are chosen so that the smallest resolvable scale (s0) is equal to 2δt (half sampling frequency) and the largest scale being δ −1 log 2 ( δ / 0 ). During CWT, the wavelet is scaled in both frequency and time domains, using a defined number of scales, Eq. (2-3). Frequency-domain scales increase exponentially from j = 0 to J (J being the Nyquist frequency). Time-domain scales are increased linearly, from n = 0 to N-1 (N equal to the length of data series). A δj value of 1/8 was chosen as a compromise between high-frequency resolution without long computational time. The average frequency-resolved coefficients over a selected segment of time give a real covariance between two signals, which in turn is used to calculate the eddy−flux (Metzger et al., 2013). Summatively, this approach provides localised highly resolved fluxes whilst accounting for all relevant transport scales.