The AOTF-based NO 2 camera

The abundance of NO2 in the boundary layer relates to air quality and pollution source monitoring. Observing the spatiotemporal distribution of NO2 above welldelimited (flue gas stacks, volcanoes, ships) or more extended sources (cities) allows for applications such as monitoring emission fluxes or studying the plume dynamic chemistry and its transport. So far, most attempts to map the NO2 field from the ground have been made with visible-light scanning grating spectrometers. Benefiting from a high retrieval accuracy, they only achieve a relatively low spatiotemporal resolution that hampers the detection of dynamic features. We present a new type of passive remote sensing instrument aiming at the measurement of the 2-D distributions of NO2 slant column densities (SCDs) with a high spatiotemporal resolution. The measurement principle has strong similarities with the popular filter-based SO2 camera as it relies on spectral images taken at wavelengths where the molecule absorption cross section is different. Contrary to the SO2 camera, the spectral selection is performed by an acoustooptical tunable filter (AOTF) capable of resolving the target molecule’s spectral features. The NO2 camera capabilities are demonstrated by imaging the NO2 abundance in the plume of a coal-fired power plant. During this experiment, the 2-D distribution of the NO2 SCD was retrieved with a temporal resolution of 3 min and a spatial sampling of 50 cm (over a 250× 250 m2 area). The detection limit was close to 5× 1016 molecules cm−2, with a maximum detected SCD of 4× 1017 molecules cm−2. Illustrating the added value of the NO2 camera measurements, the data reveal the dynamics of the NO to NO2 conversion in the early plume with an unprecedent resolution: from its release in the air, and for 100 m upwards, the observed NO2 plume concentration increased at a rate of 0.75–1.25 g s−1. In joint campaigns with SO2 cameras, the NO2 camera could also help in removing the bias introduced by the NO2 interference with the SO2 spectrum.


Introduction
Nitrogen oxides (NO x = NO + NO 2 ) play a key role in the air quality of the boundary layer. While NO is produced in combustion processes (transport, thermal power plants, etc.), NO 2 mainly appears through the reaction of NO with O 3 or HO 2 . Eventually, the photolysis of NO 2 releases an oxygen atom and a NO molecule. To a good approximation, the balance of NO and NO 2 is kept constant through this cycle of photo-chemical reactions, which substantiates the widespread use of the NO x family concept (Seinfeld and Pandis, 2006). Considering the relative ease of measuring NO 2 with visible-light spectroscopy, NO x budgets are often inferred based on NO 2 measurements and the photochemical equilibrium assumption.
The most common NO 2 remote sensing techniques rely on the differential optical absorption spectroscopy (DOAS), which is based on the fitting of radiance spectra with the effective absorption cross section of interfering species (e.g., Platt, 1994). If equipped with a 2-D sensor array, these instruments disperse the light spectrum along one dimension and record its spatial variation along the other. Building a complete hyperspectral image requires an incremental depointing of the instantaneous field of view (FOV) or a translation of the whole instrument. Typical examples of both applications can be found in Heue et al. (2008) or Lohberger et al. (2004). While the DOAS technique is well validated in terms of accuracy and sensitivity, the need for scanning the scene hampers the detection of dynamic processes. As the scene is sampled slice by slice, the final image does not show a great temporal consistency: different rows (or columns, depending on the scanning direction) are temporally disconnected from each other. The time gap can reach several minutes between both edges of the scene.
There are situations where high spatiotemporal resolution is needed. In volcanology, for instance, the so-called SO 2 cameras are now increasingly complementing the measurements performed with classical dispersive techniques like grating spectrometers (Mori and Burton, 2006;Bluth et al., 2007). Their concept is based on taking spectral images of the plume through two interference filters. One filter selects a narrow band of the incident spectrum around 310 nm, where SO 2 is still strongly absorbing, while the other one captures the light around 330 nm, where almost no more absorption takes place. The main advantages are a typical temporal resolution of 1 Hz, the capability to capture dynamic features such as puffs in the plume and the possibility to determine the plume speed from the sequence of images. The disadvantages are the interference by the plume aerosols caused by the coarse spectral resolution and the need for regular recalibration with reference cells filled with SO 2 to account for changes of illumination conditions (Kern et al., 2010). More recent concepts now use the combined information of a spectrometer with the camera spectral images (Lübcke et al., 2013), which yields a greater measurement accuracy.
We present a new instrument, a spectral imager dedicated to measuring the 2-D NO 2 field above finite sources like thermal power plants, industrial complexes, cities, volcanoes, etc. The measurement principle is close to the SO 2 camera: snapshots at two wavelengths emphasize the presence of NO 2 by taking advantage of absolute differences in the molecule absorption cross section. Contrary to the SO 2 cameras which use interference filters, the new instrument relies on an acousto-optical tunable filter (AOTF) to provide the spectral information. The AOTF can offer sufficient spectral resolution to resolve the structures of the NO 2 spectrum. The ability to discriminate between weak and strong absorption within a few nanometers virtually cuts out any sensitivity to aerosol scattering and changes of solar angles. Potential applications include urban and industrial pollution monitoring, emission fluxes estimation, satellite-product validation and volcanic plume chemistry.

Instrument concept
The AOTF-based NO 2 camera springs from the ALTIUS instrument (atmospheric limb tracker for the investigation of the upcoming stratosphere; Fussen et al., 2016). ALTIUS is a space mission project aimed at the retrieval of atmospheric species concentration profiles with a global geographical coverage and a high vertical resolution. Its primary scientific objective is to measure ozone, but NO 2 , aerosols, H 2 O, CH 4 , polar stratospheric and noctilucent clouds, and other minor species will be measured as well. Measurements will be performed in two different geometries: limb scattering and oc- Figure 1. Optical layout of the NO 2 camera seen from top. Light propagates from left to right through a pupil and a lens doublet, a polarizer selecting vertically polarized light, the AOTF, a second cross-oriented polarizer, two lens doublets and the detector.
cultations (Sun, Moon, stars, planets). To address the problem of tangent height registration of previous limb scatter instruments, a spectral imager concept based on a tunable filter has been selected. During the feasibility study, a prototype of the visible (VIS) channel (440-800 nm) was built from commercially available parts. The detailed description of this prototype is given in Dekemper et al. (2012). We will only point out the key features of the concept.
The instrument images a 6 • square FOV onto a Princeton Instrument Pixis 512B peltier-cooled CCD detector (512 × 512 pixels). The optical layout ( Fig. 1) is linear with an intermediate focal plane located close to the AOTF. To preserve the spectral homogeneity across the image, the design is made telecentric by placing an iris at the object focal point of the first lens. This ensures an identical propagation angle of all light rays through the AOTF.
The most important part of this NO 2 camera concept is the AOTF (Chang, 1974). AOTFs have been used in many areas requiring spectral images (agriculture, food industry, fluorescence spectroscopy, etc.) but received little attention from the atmospheric remote sensing community. The working principle is based on the interaction of light and sound in a birefringent crystal (see Fig. 2). By the momentum matching of the optical and acoustic waves, a narrow portion of the light spectrum is diffracted into a slightly different direction (a few degrees). If the incident radiation is linearly polarized, the diffracted beam will leave the crystal with the orthogonal polarization. The spatial and polarimetric dissociations can be combined to achieve very efficient extinction of the unwanted spectrum.
The wave vector's matching condition (Fig. 2) creates a monotonic relationship between the light wavelength and the sound frequency. The acoustic wave is launched into the crystal by a piezoelectric transducer bonded to one of its facets. Hence, selecting a particular wavelength λ simply requires us to drive the transducer to the matching frequency F (λ). The AOTF spectral transmission function (STF) closely follows a sinc 2 shape. The amplitude of the STF, which determines the filter diffraction efficiency (DE), is controlled by the acoustic power P a (λ), which also exhibits a smooth wavelength dependence. The transducer length defines the length of the acousto-optic interaction, which directly affects the AOTF bandwidth: a short transducer will induce a larger passband and vice versa.
The parameters of an AOTF are defined by the crystal elastic and optical properties and by the propagation directions of light and sound in the frame of the crystal axes (Voloshinov et al., 2007). The AOTF we used was manufactured out of a TeO 2 crystal by the company Gooch & Housego (UK). It offers an aperture of 10 × 10 mm 2 and a tuning range covering the visible spectrum. Laboratory characterization revealed a transparency better than 90 % and a DE better than 95 %. In the relevant spectral range for NO 2 measurements, i.e., around 450 nm, the STF showed a bandwidth of 0.6 nm. Typical driving frequencies were around 130 MHz, and less than 100 mW of acoustic power was needed in any circumstances. The theoretical number of resolvable spots at 450 nm is about 350 in the plane of acousto-optic interaction (horizon) and 700 in the vertical direction.

Measurement principle
There are strong similarities between the measurement principles of a filter-based SO 2 camera and an AOTF-based NO 2 camera: the FOV needs to be pointed towards the target region (e.g., a plume) while making sure that the background can still be seen in some areas of the image. Two spectral images of the scene are taken: one at a wavelength λ s where there is strong absorption by the target species and another at a wavelength λ w where there is weak absorption. In each image, the signal S ij (λ) (in e − ) recorded by pixel ij looking at the plume will be normalized by the background signal S 0 (λ) in order to quantify the extinction that took place during the crossing of the plume. The optical thickness τ ij associated with the slant column density (SCD) of the target species observed in the FOV of pixel ij follows from the comparison of the normalized signals recorded at the two wavelengths.
The major difference comes from the capability of the AOTF-based NO 2 camera to resolve the fine structures of the absorption cross section σ NO 2 (Fig. 3). This allows choosing λ s and λ w very close to each other (a few nm), minimizing the interference by broadband absorbing and scattering species like aerosols.

Mathematical model
As AOTFs do not treat different polarizations identically, an AOTF-based NO 2 camera exhibits a strong polarization sensitivity. The polarization state of a stream of light is described by the Stokes vector s = (I, Q, U, V ) T , where I = I h + I v and Q = I h − I v , with I h and I v being the light intensity along the horizontal and vertical axes of a scene frame. U and V also refer to the orientation of the polarization ellipse but they will not be discussed further because they do not participate if the AOTF and its surrounding polarizers are well aligned.
When light passes through a polarizing part, its Stokes vector can be changed. A polarizing element is therefore represented by a 4 × 4 transfer matrix: the Mueller matrix M. A chain of optical elements is represented by the product of their Mueller matrices. In our design, the light passes first through a vertical linear polarizer, then the AOTF, and finally a horizontal linear polarizer. The Stokes vector representing the light leaving the second polarizer is therefore given by s = M Ph · M AOTF · M Pv · s. The Mueller matrices of the ele-E. Dekemper et al.: The AOTF-based NO 2 camera ments are as follows: where A is the amplitude of the AOTF STF (i.e., its DE, 0 ≤ A ≤ 1), η 2 t is the attenuation of the light intensity along the polarizer transmission axis, and η 2 e is the attenuation along the extinction axis. Assuming that all three elements have their transmission and extinction axes well aligned, the total Mueller matrix of the camera is simply M = η 4 t · M AOTF . As the detector only measures the total light intensity, the first element of the Stokes vector is the only meaningful quantity: Hence, in the present configuration, the NO 2 camera is only sensitive to vertically polarized component of the light.
We now have a description of the light intensity which will be measured by the detector, but we still have to account for the transmittance of the lenses (T ) and the quantum efficiency (QE) of the detector. These terms exhibit a smooth wavelength dependence. For the AOTF STF, one can use F(λ; λ c ) = A(λ c ) · G(λ − λ c ), where G is essentially a sinc 2 function. Moreover, some parameters are susceptible to vary across the FOV, yielding a pixel-to-pixel variation. This is particularly true when image planes are located close to optical surfaces (mainly the AOTF and the detector). Finally, the electronic current (in e − s −1 ) found in pixel ij when the AOTF is tuned to λ c is given by The decision to leave the smoothly varying parameters out of the integral is supported by the narrow passband of the AOTF (0.6 nm). Their product forms the instrument response at pixel ij and wavelength λ c : r ij (λ c ). The remaining integral is simply the convolution of the vertically polarized incident light intensity with the AOTF STF.
Suppose now that pixel ij is looking through an optically thin plume. NO 2 and other species will absorb or scatter photons and decrease the background light intensity I v 0 according to the Beer-Lambert law of extinction: where τ NO 2 ij denotes the plume optical thickness caused by absorption by NO 2 along the light path ending on pixel ij , and τ * ij is the effective optical thickness of all other chemical species and particles. Over the passband of the AOTF, one can consider τ * (λ) as a constant value τ * (λ c ) and replace τ NO 2 (λ) by its weighted average: As the optical thickness is defined by the product of the trace gas SCD k with its absorption cross section σ , it is clear that Under these assumptions, one can insert Eq. (5) into Eq. (4) and write for the pixel photoelectric current: In the meantime, other pixels have been looking at the unattenuated background intensity I 0 . Suppose that one of them is pixel mn. According to Eq. (4), we have Averaging all these background-looking pixels yields the reference current associated with the background intensity: with r representing the average instrument response. Dividing C ij by C 0 yields the transmittance of the plume alone: If the spectral interval between λ w and λ s is small enough that the approximation τ * (λ w ) = τ * (λ s ) holds, then the ratio of the transmittances T (λ w )/T (λ s ) is a quantity which only depends on the NO 2 content of the plume. Introducing the relative instrument response at pixel ij , ρ ij (λ) = r ij (λ)/r(λ), we find Finally, the NO 2 SCD subtended by the area of the plume observed by pixel ij follows by taking the logarithm of the Clearly, the best sensitivity is reached by maximizing the differential optical thickness when selecting λ w and λ s .

Ancillary data
Equations (11) and (12) show that the NO 2 SCD can be obtained from a combination of measurements (the detector signal), cross-section data and the knowledge of the instrument response. In the results presented below, the cross section is taken from Vandaele et al. (1998). For the ρ ij coefficients, an ad hoc method was set up to build a synthetic flat field. Taking advantage of a cloudy weather (100 % cloudiness), longexposure frames (10 s) were captured at the required wavelengths looking at zenith. The mean image obtained from tens of such frames constitutes the instrument response to a synthetic, radiometrically flat scene. This allows us to remove wavelength-dependent nonuniformities which can be relatively pronounced in, e.g., the AOTF. Determining the photoelectric current strictly proportional to the signal (i.e., C ij and C 0 ) implies that voltage offset, dark current and stray light have been subtracted from the raw data. In this respect, AOTFs offer a unique feature: one can turn them off. This is simply done by bringing the acoustic wave amplitude to 0. An image acquired in these conditions contains anything but the real signal (i.e., dark current, offset, stray light). Using D ij and D off ij to represent the raw signal of pixel ij (in digital numbers, DN) when the AOTF is turned on or off respectively, the photo-electric signal is given by where γ is the sensor gain (in DN/e − ). The only precaution is to take these dark images regularly because the stray light is a function of the general illumination conditions (e.g., solar angles) and it will vary with local time.

Data averaging and multiple image doublets
It is often necessary to repeat the measurements in order to average out transient features and increase the signal-tonoise ratio. Assuming that only the plume optical transmission varies, we can write a time-dependent version of Eq. (7): The time-averaged optical thickness τ (λ, t) can be obtained from the geometric mean of the consecutive images: Another means of increasing the reliability of the measurements is to use different doublets, i.e., pairs of λ w and λ s . If the transmittance is known for several doublets, their product strengthens the NO 2 SCD retrieval by providing information from multiple spectral regions. If σ NO 2 = σ NO 2 (λ s ) − σ NO 2 (λ w ), then for two doublets we have for the SCD This approach can potentially attenuate a bias in one of the measurements.

Error budget and instrument sensitivity
One can work out Eq. (12) with the classical first-order Taylor expansion approximation to determine the uncertainty on the NO 2 SCD. This approach will require estimates of the uncertainty on the photon counts C ij , on the background signal C 0 , on the relative instrument response ρ ij and on the cross-section data σ NO 2 . These estimates are not always easily obtained, and we briefly discuss each of them. The photo-electric counting rates C ij are obtained from Eq. (13): C ij = S ij /t, where t is the sensor exposure time. It is reasonable to assume that the camera operator selects acquisition settings to ensure that the signal is well into the shot noise regime: σ C ij = S ij /t. With signals exceeding 10 4 e − in 1 s (the case in the examples below), the relative uncertainty on C ij will be below 1 %.
The background signal C 0 is estimated by averaging the pixels looking at the background of the scene. While one would presume that the averaging of a large number of such pixels should yield a very high precision, the accuracy is limited by the difficulty of identifying pixels effectively looking at the background. Automated data processing needs a screening of each image to determine if a pixel is looking at the plume, the background, a cloud or even a bird. This screening is based on the interpretation of the raw signals and, for instance, it sometimes fails to recognize pixels which still have in their FOV the residual NO 2 molecules left by a past position of the plume. From our experience, the relative uncertainty on C 0 determined from a single image is generally larger than 1 % (determined from the sample standard deviation). Using multiple images, as explained in Sect. 3.3, reduces this uncertainty, as C 0 is computed for each image and then averaged. A 1 % total relative error is achievable with a few images.
The relative instrument response nonuniformity ρ ij can be obtained from a homogeneous scene (i.e., a flat field such that I ij (λ) = I (λ) ∀i, j ). In this particular case, ρ ij (λ) = C ij (λ)/C(λ), where C(λ) is the average of C ij over a large number of pixels. If the flat field is built from a number of relatively homogeneous images under the assumption that their average is truly flat, then the uncertainty on the flatness participates to the error budget of ρ ij (λ) and quickly becomes the driver (signal shot noise is surpassed). This error source is a generic problem of all imaging systems but remains difficult to quantify. The only certainty is that it drops with the sample size.
The NO 2 absorption cross-section data are taken from Vandaele et al. (1998), who report a total relative uncertainty of 3 % at a resolution of 2 cm −1 (0.04 nm at 450 nm). Taking our coarser resolution into account (about 0.6 nm), the uncertainty drops to about 0.8 % for the convolved spectrum. However, the AOTF tuning curve is temperature dependent, with a typical drift of +0.1 nm per Kelvin (Ohmachi and Uchida, 1970;Uchida, 1971). The driving electronics is currently not enslaved to a temperature sensor. The exact measurement wavelength is computed at the processing stage. Depending on the amount of wavelength drift, the uncertainty on σ NO 2 (λ) can reach 5-10 %.
The minimum relative uncertainty on the NO 2 SCD will be reached if the uncertainty on the plume transmittance T is driven by C 0 . Assuming σ T /T = 1 %, and taking into account a 5 % error on the cross-section term (with a typical value for σ NO 2 (λ s )−σ NO 2 (λ w ) = 2×10 −19 ), one obtains σ k = 5 × 10 16 molecules cm −2 . If one assumes less favorable conditions like a 1 % uncertainty on ρ, yielding σ T /T = 2 % and a 10 % error on the cross section, then the SCD error reaches 10 17 molecules cm −2 .
4 Application to the remote sensing of NO 2 at a coal-fired power plant The data of a spectral imager such as the NO 2 camera are more easily exploited if a number of observational requirements are satisfied. First, the camera must be placed at a location where both the plume and the background can be captured within the same image. Second, the target plume must remain optically thin in order to preserve the assumption of the Beer-Lambert extinction along a straight light path. Finally, scattered clouds behind the plume will corrupt the retrieval and should be avoided. These three requirements were sometimes fulfilled during the second Airborne ROmanian Measurements of Aerosols and Trace gases (AROMAT-2) campaign in August 2015. The campaign aimed at joining the efforts of several Euro- Figure 4. Observational geometry during the AROMAT-2 campaign at Turceni's power plant. The NO 2 camera was installed on a football pitch looking at the four 280 m tall stacks. The red lines delimit the camera horizontal FOV (6 • ). The direction of the Sun at 16:00 local time is approximately indicated, together with two rays illustrating the scattering behind the scene towards the camera. One of the rays passes through the plume, while the other one passes by. Map data from OpenStreetMap. pean research institutes and universities to spatially and temporally characterize the emissions from two types of sites: a large city (Bucharest) and point sources (large thermal power plants in the Jiu Valley, Romania). Both sites should eventually serve as validation targets for the ESA TROPOMI/S-5P mission.
The NO 2 camera was placed at a distance of 2.5 km from a group of four stacks belonging to Turceni's power plant, the largest being in Romania (330 MW per turbine, 2000 GWh year −1 total electric power generation of which more than 93 % is generated from coal). Figure 4 depicts the measurement geometry. Our location was 44.6792 • N, 23.3788 • E, the line of sight (LOS) azimuth angle ranged from 113 • (left edge of the image) to 119 • (right edge) eastward from north, and the LOS zenith angle ranged between 75.5 • (top edge) and 81.5 • (bottom edge). We only report on measurements performed on 24 August between 16:15 and 16:30 LT as the observational conditions were close to ideal and best illustrate the performance of the instrument. In particular, the smokes were optically thin, with the blue sky clearly visible in the background. This ensures that absorption is the dominant process over scattering for the extinction of light rays crossing the plumes (Beer-Lambert regime). The optical thickness of the smokes was always smaller than 0.1 at our measurement wavelengths.

Exhaust plume NO 2 SCD field
As explained in Sect. 3.1, the 2-D NO 2 SCD field is computed from at least two spectral images recorded at wavelengths showing a significant difference of absorption cross section. To increase the reliability of the measurements, four doublets of wavelengths were used: λ w1 = 441.8 and λ s1 = 439.3; λ w2 = 446.7 and λ s2 = 448.1; λ w3 = 437.9 and λ s3 = 435.1; λ w4 = 465.8 and λ s4 = 463.2. The automated acquisition system was in charge of synchronizing the driving of the AOTF with the image acquisition. A nominal acquisition sequence started by setting the appropriate acoustic signal for the AOTF to filter at λ w1 , opening the CCD shutter for 0.5 s, reading out the image and repeating these operations for the seven other wavelengths. After completion of the nominal sequence, a picture with the AOTF turned off is taken and the nominal sequence is resumed. The dwell time between the closing of the shutter and its reopening was 1.3 s, yielding a total acquisition sequence duration of 13.1 s for the 8 spectral images. In the plane of the stacks, the image footprint spans an area of 250 × 250 m 2 with a 50 cm sampling.
The data analysis revealed that the images from the second and fourth doublets were the less noisy because of a larger natural radiance and sensor sensitivity compared to the wavelengths of doublets 1 and 3. Also, due to the plume displacement over time (wind) and the presence of moving and changing inhomogeneities across the plume (puffs, turbulent eddies), it was necessary to perform time averaging (Sect. 3.3). Indeed, the 1.3 s between two consecutive images is already a long time for features moving at a typical 5 m s −1 speed (corresponding to 10 pixels per second). Figure 5 shows the NO 2 SCD field retrieved from the averaging of images taken at λ w2 , λ s2 , λ w4 and λ s4 (12 of each) using the method described in Sect. 3.3. For each wavelength, the background signal C 0 was determined from image areas unaffected by the plume. The relative error on C 0 is about 0.5 % (estimated from the standard deviation σ C 0 of the pixels sample yielding C 0 ). Within this precision, no variation of C 0 across the FOV could be significantly detected. The reason is the relatively small FOV of the camera (about 6 • ) combined with a high Sun at the time of the measurements (making the scene illumination quite homogeneous). In Fig. 5, the background grayscale image is the mean image at λ w4 , whereas the pixels where the SCD is computed were selected based on the criterium C ij < C 0 −2σ C 0 . Investigating the random fluctuations observed in various areas of the SCD field, one can estimate the detection limit to about 5 × 10 16 molecules cm −2 .

NO 2 emission fluxes and synergies with SO 2 cameras
The capability of resolving the NO 2 SCD field with a high spatial and temporal resolution provides new possibilities for the understanding of the plume chemistry. Coal combustion processes yielding the formation of nitrogen and sulfur species are well known (Flagan and Seinfeld, 1988), and several reactive plume models can simulate the transport, formation and removal of these species over different scales. These models are generally validated by in situ air sampling at distances of several kilometers downwind (see for instance Chowdhury et al., 2015). Very few experiments attempted to characterize the reactive content of the early plume, where the reactions are still governed by the combustion products (Hewitt, 2001). In most cases, a DOAS scanning system was used (Lee et al., 2014(Lee et al., , 2009Lohberger et al., 2004). The same technique was also used for SO 2 , but to a lesser extent since the introduction of filter-based SO 2 cameras (Smekens et al., 2015). Recently, imaging Fourier transform spectroscopy (IFTS) demonstrated capability for the measurement of a number of mid-infrared emitting species such as CO 2 and SO 2 (Gross et al., 2010). However, NO, but not NO 2 , can be retrieved with this technique. An undisputed advantage of imaging systems with high temporal resolution is their ability to track the displacement of remarkable features from one image to another. We used the complete time series of spectral images (50 sequences of 8 spectral images at a rate of 0.5 Hz) to determine the vertical speed of the plume. This was done by tracking signal features created by local increase or decrease of the NO 2 concentration. On average, a vertical speed of 4.8 ± 0.5 m s −1 was ob- served. Furthermore, assuming a Gaussian dispersion of the plume, one can infer a circular cross section from the apparent width of the plume at each detector row (i.e., every 50 cm above the stack outlet). As a result, a profile of emission flux (in g s −1 ) can be drawn. Figure 6 shows the NO 2 emission flux as a function of altitude up to a height above which the two plumes cannot be discriminated anymore. The fluxes were calculated from the two SCD maps of Fig. 5 and both stacks. The increase is the result of the conversion of NO into NO 2 mainly by the reactions 2NO + O 2 → 2NO 2 and NO + HO 2 → NO 2 + OH (Flagan and Seinfeld, 1988;Miller and Bowman, 1989), even if these processes are balanced by the photodissociation of NO 2 as soon as it reaches open air under daylight (NO 2 +hν → NO+O). Qualitatively, these results agree well with the increase reported by Lee et al. (2009) in a study of the rate of increase of NO 2 above power plant stacks. The analysis of Fig. 6 reveals that within the method approximations, the NO 2 concentration in the plume increases at a rate ranging from 0.75 to 1.25 g s −1 (9.8×10 21 -1.6×10 22 molecules s −1 ) on average for the first 20 s.
The knowledge of the spatial distribution of NO 2 can also prove useful to correct measurements marked by interference from NO 2 . A good example is with SO 2 cameras where the SO 2 SCD field is retrieved by comparing the plume transmittance around 310 and 330 nm. In this range, NO 2 is also absorbing and its cross section roughly doubles from 310 to 330 nm. Therefore, if both molecules are present in the plume, the SO 2 camera alone cannot distinguish their respective signatures. So far, this interference has been over-looked in SO 2 camera validation exercises (Smekens et al., 2015;Kern et al., 2010). In the case of the plumes shown in Fig. 5 for instance, a SO 2 camera such as the one used by Smekens et al. (2015) would observe a τ NO 2 = 0.04 when the NO 2 SCD reaches 3 × 10 17 molecules cm −2 . This variation of optical thickness corresponds to a SO 2 SCD of about 1.6×10 17 molecules cm −2 , which is twice the detection limit reported in Smekens et al. (2015). Clearly, the bias would increase with higher concentrations of NO 2 . Taking advantage of the similar spatial resolution of both instruments, the NO 2 camera can provide a complete correction map for the SO 2 data. On the temporal resolution side, however, the NO 2 camera is, at the moment, not capable of following the pace of SO 2 cameras (1 Hz typical), such that the correction maps would have to be applied to temporally averaged SO 2 data.

Conclusions
We have described a new passive atmospheric remote sensing instrument for the measurement of NO 2 SCDs above strong sources. It is based on an AOTF which offers a sufficient acceptance angle to be placed in an imaging system and the necessary resolution for taking advantage of the fine structures of the NO 2 absorption cross section. The AOTF is electrically driven, such that fast synchronized acquisitions of spectral images are possible.
The measurement principle is similar to the filter-based SO 2 camera: SCDs are retrieved from at least two spectral images taken at wavelengths where absorption by the target molecule is significantly different. Wavelengths are picked in the range 440-470 nm. Thanks to its higher spectral resolution, the AOTF-based NO 2 camera can perform its measurements within a few nanometers. This makes the sensitivity to aerosols negligibly small.
A mathematical framework for data processing has been developed, and the different sources of error have been addressed. In applications focusing on relatively high spatiotemporal resolution, the NO 2 SCD detection limit is about 5×10 16 molecules cm −2 . Different measurement geometries offering longer integration times or more stable targets would yield a lower limit.
The NO 2 camera was successfully tested during the AROMAT-2 campaign where measurements of NO 2 SCD fields above the flue gas stacks of a coal-fired power plant were performed with a temporal resolution of 3 min and a spatial sampling of 50 cm (for a complete scene of 250 × 250 m 2 ). Values up to 4 × 10 17 molecules cm −2 were observed. The quality of the data allowed us to clearly identify the conversion process from NO to NO 2 in the early plume, providing quantitative information on the plume dynamic chemistry. In another example of application, the measurements were used to show how the knowledge of the highresolution NO 2 field can help to correct SO 2 camera data. If overlooked, the interfering absorption of NO 2 can yield a significant bias in the retrieved SO 2 SCDs. Other applications range from emission monitoring to volcanic plume chemistry.
While the concept is mature, a number of improvement directions are still being investigated. The most promising ones are the implementation of a temperature feedback loop to reduce the uncertainty on the filtered wavelength and the replacement of the CCD by a CMOS in order to reduce the cooling needs and increase the temporal resolution of the measurements.

Data availability
The data are available upon request to the contact author.
Author contributions. Emmanuel Dekemper developed the NO 2 camera measurement principle, led the characterization and the participation in the AROMAT-2 campaign, and processed the data. Bert Van Opstal and Jurgen Vanhamel developed the acquisition software and the AOTF driving electronics and participated in the campaign. Didier Fussen is at the origin of the instrument concept and supported its development in the frame of the ALTIUS project. The authors declare that they have no conflict of interest.