2.1 MAX-DOAS measurements

The MAX-DOAS instrument is set up on the platform on the fifth floor of the UFS (47.417^∘ N, 10.980^∘ E), about 20 m above ground level, which is about 2650 m above sea level. The instrument consists of a scanning telescope, a stepping motor which controls the viewing zenith angle of the telescope, and two spectrometers covering the ultraviolet (UV) and visible (VIS) wavelength bands. Incoming sunlight is redirected by a prism reflector and a quartz fiber bundle to the spectrometers for spectral analysis. The field of view (FOV) of the instrument is ∼0.98^∘. Two spectrometers (OMT Instruments, OMT ctf-60) each equipped with a charge-coupled device (CCD) detector are used to measure the spectra of both UV (320–478 nm) and VIS (427–649 nm) wavelength ranges. The full width at half maximum (FWHM) spectral resolutions of the UV and VIS spectrometers are about 1.0 and 0.6 nm, respectively. The scanning direction of the telescope is controlled by the stepping motor.

As the measurement geometry is limited by the topography, the viewing azimuth angle of the telescope was adjusted to due south (180^∘) with the lowest elevation angle of 1^∘. Each scanning cycle consists of measurements at elevation angles (α) of 90^∘ (zenith), 30, 20, 10, 5, 2, and 1^∘. A single measurement at each elevation angle lasts for ∼1 min, and a full scanning cycle takes about 10 min. The recorded spectrum of each measurement is the sum of the CCD readouts within ∼1 min. In order to optimize the measurement SNR, avoid saturation, and achieve a constant signal level, the data acquisition software automatically adjusts the exposure time of each readout to make the maximum count close to 70 % of saturation level (65 535 counts). Depending on the intensity of received light, the exposure time of each readout varies from tens of milliseconds to a few seconds. The measurements of the UV and VIS bands are taken by the two spectrometers simultaneously, but their exposure times are adjusted individually. The instrument takes measurements continuously during daytime (solar zenith angle (SZA) <85^∘), but during the noon ($\mathrm{175}{}^{\circ }<$ solar azimuth angle (SAA) <185^∘) and twilight periods ($\mathrm{85}{}^{\circ }<$ SZA <92^∘), the instrument takes only zenith measurements.

The MAX-DOAS instrument has been running since February 2012. However, the measurement was interrupted between February 2013 and July 2013 due to instrument maintenance. In February 2016, the measurement was interrupted again, and the VIS spectrometer was found to be degraded. In this paper, we present 4 years of MAX-DOAS measurements from February 2012 to February 2016.

2.2 Sun photometer measurements

Next to the MAX-DOAS instrument, a sun photometer is installed at the UFS, which provides measurements of radiances at 12 wavelengths between 340 and 1640 nm with a temporal resolution of 1 s. The instrument was developed at the Meteorological Institute of Ludwig Maximilian University of Munich (LMU) based on a system operated in the framework of the SAMUM campaigns (Toledano et al., 2009, 2011) but with improved electronics and data acquisition developed by Physikalische Messsysteme Ltd. In this study, the AODs derived from sun photometer measurements applying the well-established Rayleigh calibration method were used for the intercomparison with the MAX-DOAS retrieval. For this purpose, AOD measurements at 340 and 380 nm were interpolated to 360 nm while AODs at 477 nm were interpolated from the measurements at 440 and 500 nm. The interpolation followed the Ångström exponent method. Measurements were given as hourly averages. Due to the reduced accuracy under large SZA, only the measurements between 10:00 and 14:00 UTC each day were used. In order to ensure the data quality, only cloud-free conditions and periods of stable aerosol abundance (variability of radiances below 5 % within 1 h) were considered. These requirements reduce the number of available sun photometer measurements considerably. Note that the AOD is often below 0.02 at the relevant wavelengths with an uncertainty on the order of ±0.015 due to calibration errors, Rayleigh correction and radiometric accuracy. As the uncertainty of the AOD measured by the sun photometer is relatively large, the uncertainty of the Ångström exponent would be further amplified. Consequently they are not used in this study.

The aerosol optical properties which are not available from the UFS measurements but required for our MAX-DOAS inversion scheme (single-scattering albedo and phase function) were estimated from the AERONET measurements at Hohenpeißenberg, which is located at an altitude of 980 m and approximately 43 km north of the UFS. The AERONET data were available at 440, 675, 870, and 1020 nm. Therefore, the data at 360 nm were extrapolated, and the data at 477 nm were interpolated. As Hohenpeißenberg and UFS are located at different altitudes, the aerosol optical properties might be slightly different. Therefore, we have analyzed the uncertainties caused by the differences in single-scattering albedo and phase functions through a sensitivity analysis. The results show that the influences of aerosol optical properties are in general less than 3 %; see Appendix B3. Some other MAX-DOAS studies also found that aerosol optical properties show only small impacts on aerosol profile retrieval (e.g., Chan et al., 2019).

2.3 Ceilometer measurements

The UFS is also equipped with a Lufft (previously Jenoptik) ceilometer (model CHM15kx; see Wiegner and Geiß, 2012) operated by the German Weather Service (DWD). Ceilometers are single-wavelength backscatter lidars, and the received signals follow the well-known lidar equation (Wiegner et al., 2014). The CHM15kx is eye-safe and fully automated, which allows unattended 24/7 operation. It can be used to monitor aerosol layers (e.g., volcanic ash; see Schäfer et al., 2011) and validate meteorological and chemistry transport models (see, e.g., Emeis et al., 2011) and is foreseen for model assimilation (e.g., Wang et al., 2014b; Warren et al., 2018; Chan et al., 2018).

The CHM15kx ceilometer is equipped with a diode-pumped Nd:YAG laser emitting pulses at 1064 nm. The received backscatter signals are stored in 1024 range bins with a resolution of 15 m. The temporal resolution is set to 15 s. The signals are corrected for incomplete overlap by a correction function provided by the manufacturer.

Figure 1Seasonal average aerosol extinction profiles extracted from ceilometer measurements.

Download

A strict retrieval of the particle extinction coefficient from ceilometer measurements is not possible due to the unknown lidar ratio; furthermore, exploitation of the signal in the range of incomplete overlap is subject to errors. Thus, in order to convert the ceilometer measurements to aerosol extinction profiles, we followed an approach mentioned in Wagner et al. (2019). The range-corrected attenuated backscatter data from July 2016 to December 2017 were seasonally averaged. Data of the altitude between 500 m and 5 km above the instrument were averaged with a vertical grid resolution of 500 m. Data below 500 m were assumed to be constant, following the values at 500 m. The extinction coefficients were first calculated by scaling the attenuated backscatter profiles (β^∗) to the seasonal average AODs at 360 and 477 nm obtained from the sun photometer. The extinction profiles were then used to correct for the attenuation of the backscatter profiles following the lidar equation (Klett, 1981; Fernald, 1984). The corrected backscatter profiles (β) were then scaled to the AODs at 360 and 477 nm measured by the sun photometer to obtain the extinction profiles; see Fig. 1. Note that the ceilometer measures at 1064 nm and the optical properties of aerosols depend on the wavelength. Therefore, the uncertainties of these profiles are very large and they should be considered qualitative only.

The results shown in Fig. 1 indicate that the aerosol load at the UFS is highest in summer (June, July, and August) and lowest in winter (December, January, and February). The seasonal results also indicate large variations in the aerosol load from the surface up to 2 km. Above 2 km the variability is smaller; however, their contribution to the total column is still substantial (∼30 %–50 %).

3.1 O₄ DSCD calculation

The DSCDs of O₄ were derived from both UV and VIS spectra using the DOAS technique (Platt et al., 1979; Platt and Stutz, 2008). In the retrieval, DSCD is defined as the difference between the slant column density (SCD) of each off-zenith spectrum ($\mathit{\alpha }\ne \mathrm{90}{}^{\circ }$) and the corresponding zenith reference spectrum ($\mathit{\alpha }=\mathrm{90}{}^{\circ }$). The QDOAS spectrum analysis software (version 3.2) developed by BIRA-IASB (http://uv-vis.aeronomie.be/software/QDOAS/, last access: 1 April 2020) was used for the spectral fitting analysis. The calibration of the spectrometers was performed by fitting the measured solar spectra to the literature solar reference (Chance and Kurucz, 2010). All the measured spectra were first corrected for offset and dark current.

Details of the DOAS fit settings for both bands are listed in Table 1. The fitting windows were determined according to both the absorption signal of O₄ and the SNR of the spectrometers. For UV spectra, the fitting window is 338–370 nm, which is the same for most of the other MAX-DOAS studies (e.g., Clémer et al., 2010; Wang et al., 2014a; Kreher et al., 2019), and it covers the strong absorption peak at 360.8 nm and a weak absorption peak at 344 nm. For VIS spectra, because the spectral range of the spectrometer begins at 427 nm and the SNR close to the spectral edges is low, we therefore adapted a smaller fitting window of 440–490 nm, which is a bit narrower than the fitting window of 425–490 nm commonly used in other MAX-DOAS studies (e.g., Clémer et al., 2010; Chan et al., 2017; Kreher et al., 2019). The VIS fitting window covers the strong absorption peak at 477 nm and a weak absorption peak at 446.5 nm. As the temperature at the UFS typically varies between 263 and 279 K (Risius et al., 2015), trace gas absorption cross sections measured at 273 K were used in the DOAS fit. Absorption cross sections of several trace gases as well as a synthetic ring spectrum were included in the DOAS fit. For each scanning cycle, the zenith spectra before and after the cycle were temporally interpolated to the measurement time of each off-zenith spectrum. The broadband spectral structures caused by Rayleigh and Mie scattering were removed by including a low-order polynomial in the DOAS fit. Small shift and squeeze of the wavelengths were allowed in the wavelength mapping process in order to compensate for small uncertainties caused by the instability of the spectrograph.

The root mean square (rms) of fit residual was used to evaluate the performance of the DOAS fit. DSCDs with a residual rms larger than $\mathrm{1}\times {\mathrm{10}}^{-\mathrm{3}}$ were not considered in the following analysis. Under cloud-free conditions, the residual rms of most of the UV spectra varies between $\mathrm{5}\times {\mathrm{10}}^{-\mathrm{4}}$ and $\mathrm{9}\times {\mathrm{10}}^{-\mathrm{4}}$, while the residual rms of most of the VIS spectra varies between $\mathrm{2}\times {\mathrm{10}}^{-\mathrm{4}}$ and $\mathrm{5}\times {\mathrm{10}}^{-\mathrm{4}}$. This is because both the light intensity and the O₄ absorption are stronger at the VIS band; hence the measurement SNR is higher.

Table 1The DOAS fit settings for the UV (338–370 nm) and VIS (440–490 nm) bands.

^a I₀ correction is applied with SCD of 10¹⁷ molec. cm⁻² (Aliwell et al., 2002). ^b I₀ correction is applied with SCD of 10²⁰ molec. cm⁻² (Aliwell et al., 2002).

Download Print Version | Download XLSX

3.2 Cloud screening

The aerosol profile retrieval requires the forward simulation of the radiative transfer in the atmosphere. As the radiative transfer is rather complicated for cloudy-sky conditions, the forward simulation usually assumes a cloud-free atmosphere. The aerosol retrieval might result in large uncertainty under cloudy or foggy conditions. Therefore, it is important to filter out the measurements taken under cloudy or foggy conditions. In this study, a cloud screening approach based on a color index (CI) (Wagner et al., 2014, 2016) was applied to filter out cloudy measurements. The CI is defined as the ratio of radiative intensities at 330 and 390 nm in this study. Larger CI indicates the UV–VIS intensity ratio is higher; hence, the sky is more blue. Our cloud screening method is presented in Appendix A. The cloud screening results during the entire measurement period are summarized in Table 2. Among the four seasons, the percentage of cloudy measurements is highest in summer and lowest in winter. In total, about 60 % of the zenith measurements were determined as cloudy scenes, and the corresponding scanning cycles were not used in the aerosol profile retrieval.

Table 2Summary of cloud screening results.

Download Print Version | Download XLSX

Figure 2Altitude and type of the ground surface under the viewing direction (due south) of the MAX-DOAS at the UFS. Both the altitude data and surface type are obtained from Google Earth. The shadow of the line of the 1^∘ viewing direction indicates the FOV of the telescope which is about 0.98^∘.

Download

3.3 Topography effect and the simplification in radiative transfer model

The topography around the UFS is quite complex, which complicates the radiative transfer simulations. As shown in Fig. 2, the surface altitude varies between 600 and 2800 m a.s.l. along the viewing direction of the MAX-DOAS instrument. Figure 2 also shows the type of surface in different colors which includes forests, meadows, rocks, etc. Some parts of the surface are seasonally or permanently covered by snow, while some steep slopes cannot be covered by snow even in winter.

Three-dimensional radiative transfer models (RTMs) can consider such a complex terrain, but they are computationally expensive and unaffordable for retrieval. Due to the limitation of the two-dimensional RTM LIDORT (Spurr et al., 2001; Spurr, 2008) used in the study, we simplified the ground topography to a flat surface at an altitude of 2650 m a.s.l. in the radiative transfer simulations. In order to estimate the error caused by this simplification, we investigated using the three-dimensional RTM TRACY-2.

TRACY-2 is a full spherical Monte Carlo atmospheric RTM (Deutschmann, 2008; Wagner et al., 2007), which allows the simulation of three-dimensional radiative transport as well as two-dimensional variation in the surface height. The model was compared to other RTMs and very good agreement was found (Wagner et al., 2007). We also performed an intercomparison with LIDORT. The result shows that with the same definition of topography and atmosphere, the difference between the O₄ DSCDs simulated by the two RTMs is less than 3 %.

For the three-dimensional simulations carried out in this study, a pseudo-reality topography was defined with the exact ground altitude (obtained from Google Earth) in the azimuth direction of the MAX-DOAS measurements taken into account, whereas in the dimension orthogonal to this direction the surface altitude was set constant. This simplification was chosen to reduce the computational effort. We feel that this approach is justified since the atmospheric light paths in the viewing direction of the instruments can be very large (up to several tens of kilometers). It is most important to take this variation in the surface altitude along this direction into account, whereas the influence of the orography perpendicular to this direction is expected to be small.

Figure 3Relative differences of O₄ DSCDs at (a, c) 360 nm and (b, d) 477 nm simulated with a flat surface at 2650 m compared to the O₄ DSCDs simulated with the pseudo-reality topography. Panels (a) and (b) show the results simulated with the same SZA of 50^∘ and different RAAs (relative solar azimuth angles) of 30, 60, and 90^∘. Panels (c) and (d) show the results simulated with the same RAA of 60^∘ and different SZAs of 30, 50, and 70^∘. Solid curves are the results simulated under aerosol-free conditions, and dashed curves are the results simulated with a box-shape profile with AOD =0.12 and box height =3 km.

Download

Simulations were performed with all the combinations of three different SZAs (30, 50, and 70^∘), three different relative solar azimuth angles (RAAs) (30, 60, and 90^∘), and two different aerosol extinction profiles (an aerosol-free profile and a box-shape profile with AOD =0.12 and box height =3 km), i.e., altogether 18 cases. For each case, O₄ DSCDs at 360 and 477 nm were simulated with both the flat surface at 2650 m and the pseudo-reality topography using TRACY-2. The relative errors of O₄ DSCDs simulated with the flat surface compared to those simulated with the pseudo-reality topography are calculated. A fixed surface albedo of 0.07 was used in the simulations. For both wavelengths, the single-scattering albedo was set to 0.93 and the phase function was defined as a Henyey–Greenstein phase function with the asymmetry parameter set to 0.68. The atmospheric profile was defined as the US standard midlatitude atmosphere (Anderson et al., 1986). Figure 3 shows the results of some of the cases: panels (a) and (b) show the results of six cases with SZA =50^∘ and different RAAs and both aerosol extinction profiles; panels (c) and (d) show the results of six cases with RAA =60^∘ and different SZAs and also both aerosol extinction profiles.

Table 3Systematic and random errors caused by the topography simplification. Results are calculated from the relative differences of O₄ DSCDs simulated with a flat surface at 2650 m compared to those simulated with the pseudo-reality surface in 18 cases (see text). The mean of the relative difference for each elevation angle and each wavelength is considered to be the systematic error. The standard deviation of the relative difference is considered to be the random error.

Download Print Version | Download XLSX

As shown in all the panels of Fig. 3 as well as in all the other cases not shown, O₄ DSCDs simulated with the flat surface are in general slightly underestimated compared to the pseudo-reality topography. The difference could be explained by the scattering in the valleys where the concentration of O₄ is higher. For the flat surface at 2650 m, the light paths below 2650 m would not be taken into account, and hence the O₄ DSCDs would be underestimated. Moreover, the relative error has no obvious correlation with elevation angle, SZA, RAA, and aerosol load. This is because the light path below 2650 m is influenced by the topography, and the influence differs with the observation geometry. In addition, the light path is also influenced by the aerosols both below and above 2650 m. Since only a pseudo-reality surface and a constant surface albedo are used in the study, the actual error caused by the topography simplification is expected to be much more complicated.

In order to make the compensation feasible, we consider the error to be the combination of a systematic error and a random error. Based on the results of all 18 cases of this study, the mean bias for each elevation angle and each wavelength is considered to be the systematic error, while the standard deviation of the relative difference is considered to be the random error; see Table 3. In the aerosol profile retrieval, systematic errors are first corrected from the measured O₄ DSCDs, while random errors are included in the error budget in the calculation of cost functions (see Sect. 3.7.2). In the following text, measured O₄ DSCDs refer to the values corrected by the systematic error unless otherwise mentioned.

3.4 Sensitivity analysis

In order to make full use of the measurement sensitivity and reduce unnecessary computational efforts, our retrieval algorithm was designed according to the sensitivity of O₄ absorption. We performed several sensitivity analyses to determine the optimal vertical grid, step size of the aerosol extinction for each layer, and maximum aerosol extinction. In addition, these sensitivity analyses also help to estimate the measurement and model errors, which are very important for the retrieval. The sensitivity analyses are based on forward simulations of O₄ DSCDs using LIDORT. We tested the sensitivities of O₄ absorption to surface albedo, aerosol optical properties, and the aerosol vertical profile. The results of the sensitivity analyses are shown in Appendix B. The extreme and median values of the parameters are also discussed in that section.

3.5 Parameterization of the aerosol extinction profile

As discussed in Appendix B4, O₄ absorption is insensitive to aerosols above 2 km. Therefore, our retrieval only focuses on aerosols between 0 and 2 km above the MAX-DOAS instrument (i.e., 2650–4650 m a.s.l.). In order to limit the complexity of the retrieval, avoid unreasonable results, and make full use of the measurement sensitivity, we parameterize the aerosol extinction profile as aerosol extinctions in three layers. The thicknesses of the two lower layers are defined as 0.5 km. Due to the lower sensitivity at high altitude, the thickness of the third layer is set to 1 km. The aerosol profile is denoted as a three-dimensional state vector x,

where σ₁ is the aerosol extinction coefficient between 0 and 0.5 km (2650–3150 m a.s.l.), σ₂ is the aerosol extinction coefficient between 0.5 and 1 km (3150–3650 m a.s.l.), and σ₃ is the aerosol extinction coefficient between 1 and 2 km (3650–4650 m a.s.l.). The definition of x is illustrated in Fig. 4a. The vertical resolution of our retrieval grid is lower compared to that of many other studies (e.g., Clémer et al., 2010; Chan et al., 2017; Tirpitz et al., 2020); however, the vertical gradient of aerosol extinction at such a high-altitude site is expected to be small and this is also proved by the ceilometer measurements. Therefore, the coarse resolution is considered to be sufficient for the retrieval of the UFS MAX-DOAS measurements.

Figure 4Definitions of (a) the parameterized aerosol profile (x) and (b) the profile set (X_LUT). Note that only some representative nodes are shown in panel (b).

Download

In order to formulate the lookup table, we defined a profile set (denote as X_LUT) which is assumed to include all possible aerosol extinction profiles under cloud-free conditions. X_LUT is a finite set of x, and the variation steps of σ₁, σ₂, and σ₃ were determined according to the sensitivity and accuracy of measurement. X_LUT includes only the profiles with reasonable shapes, and the variation range of σ₁, σ₂, and σ₃ covers the actual aerosol load at the UFS. In this way, unreasonable and unrealistic retrieval results can be avoided.

As discussed in Appendix B6, the measurement sensitivity decreases with increasing surface aerosol extinction, and the sensitivity is very low when the surface aerosol extinction coefficient exceeds 0.3 km⁻¹. Therefore, σ₁ is defined to vary between 0 and 0.3 km⁻¹. The variation step increases from 0.001 km⁻¹ per step to 0.02 km⁻¹ per step with increasing aerosol extinction, so that the difference of O₄ DSCD per step is similar to the average spectral fitting error (∼2 %). In total, we define 65 values for σ₁; see Table 4.

As illustrated in Fig. 4b, the values of σ₂ and σ₃ are defined as a tree, which means we define different values of σ₂ for each σ₁, and the values of σ₃ are also defined depending on σ₂. According to the ceilometer observations at the UFS, strong elevated aerosol layers are unlikely to exist under cloud-free conditions. Therefore we allow only weak elevated layers in designing the profile set. We assume that for reasonable profiles, σ₂ should not exceed σ₁ by more than 30 %, and σ₃ should not exceed σ₂ by more than 30 %, either. According to the sensitivity, for each value of σ₁ (σ₁>0), we define 14 possible values for σ₂ which varies from 0 to 1.3σ₁ with a step size of 0.1σ₁. In the case σ₁=0, elevated layers are not considered, and then σ₂ and σ₃ can only be 0. Similarly, σ₃ varies between 0 and 1.3σ₂. Due to the lower measurement sensitivity at high altitude, we define nine possible ratios between σ₃ and σ₂ (see Table 4). In the case σ₂=0, σ₃ can only be 0.

X_LUT includes the profiles with all the combinations of σ₁, $\frac{{\mathit{\sigma }}_{\mathrm{2}}}{{\mathit{\sigma }}_{\mathrm{1}}}$, and $\frac{{\mathit{\sigma }}_{\mathrm{3}}}{{\mathit{\sigma }}_{\mathrm{2}}}$. For each of the 64 nonzero values of σ₁, there are $\mathrm{1}+(\mathrm{13}\times \mathrm{9})=\mathrm{118}$ corresponding profiles. For σ₁=0, there is only one profile with ${\mathit{\sigma }}_{\mathrm{1}}={\mathit{\sigma }}_{\mathrm{2}}={\mathit{\sigma }}_{\mathrm{3}}=\mathrm{0}$. Therefore, the profile set consists of $\mathrm{1}+\mathrm{64}\times \mathrm{118}=\mathrm{7553}$ aerosol extinction profiles in total.

Table 4Definition of input parameters of the O₄ DSCD lookup table.

Download Print Version | Download XLSX

3.6 Definitions of other dimensions of the lookup table

The basic idea of the lookup table method is to replace the repetitive time-consuming computation with a pre-calculated array. In this study, we replace the forward simulation of O₄ DSCDs with a lookup table, so that all the possible aerosol extinction profiles can be considered in the retrieval of each measurement cycle with an affordable computational effort.

Besides the parameterized aerosol extinction profile x, we consider another four input parameters for the forward simulation which can be described as a function,

where ΔS_s refers to the simulated O₄ DSCD, λ represents the wavelength, α indicates the elevation angle, θ is the SZA, and ϕ is the RAA. All the input parameters are well-known in the retrieval.

In order to formulate the lookup table, the input parameters need to be parameterized as a grid with finite nodes. As already presented in Sect. 3.5, the aerosol extinction profile (x) is parameterized as a profile set which consists of 7553 possible profiles. As the simulated O₄ DSCDs are used to fit to the measured ones, only the data at 360 and 477 nm and at the six non-zenith elevation angles of the measurement cycles are included in the lookup table. SZA (θ) and RAA (ϕ) are parameterized as a grid with $\mathrm{1}{}^{\circ }\times \mathrm{1}{}^{\circ }$ resolution. The grid includes 5005 combinations of SZA and RAA, which can cover all possible solar positions for the daytime measurements at the UFS. When we obtain data from the lookup table, as the input SZA and RAA are not integers, the output ΔS_s is interpolated from the data of the four adjacent nodes of the SZA–RAA grid. In total, the five input parameters are parameterized as a grid with $\mathrm{7553}\times \mathrm{2}\times \mathrm{6}\times \mathrm{5005}=\mathrm{453}\mathrm{633}\mathrm{180}$ nodes. Details of the parameterization of the input parameters are summarized in Table 4.

As discussed in Appendix B, besides the input parameters we defined, O₄ DSCDs can also be affected by other parameters such as the ground albedo, aerosol optical properties, and others. Since accurate measurements of these parameters are not available and their influence is relatively small, they are considered uncertainties. In creating the lookup table, these parameters were fixed to the median values. Details of the simulation settings are listed in Table 5. O₄ DSCDs corresponding to all the nodes of the lookup table were simulated using LIDORT.

As discussed in Appendix B5, the influence from the aerosols above 2 km is also considered a kind of uncertainty and treated in a similar way as the other unknown parameters. In the simulations for creating the lookup table, the aerosol extinction coefficient between 2 and 4 km was defined as 0.5σ₃, so that this so-called parameter is fixed to the “median” value. Note that the aerosol extinction coefficient above 2 km is neither considered a part of the retrieved profile nor counted in the retrieved AOD.

Table 5Settings of fixed parameters in calculating the O₄ DSCD lookup table.

Download Print Version | Download XLSX

3.7 Error estimation

Most of the other MAX-DOAS studies only consider the spectral fitting error in their retrieval. However, this fitting error only contributes to a small part of the total error. In addition, the total error is not directly proportional to the spectral fitting error. As the measurement and simulation uncertainties play an important part in our inversion method, we perform a comprehensive error analysis for the MAX-DOAS measurement and radiative transfer simulation of O₄ DSCDs. In this study, errors from seven major sources are taken into account in estimating the total uncertainty.

3.7.3 Total uncertainty

We assume that the seven kinds of errors mentioned in Sect. 3.7.1 and 3.7.2 follow the normal distribution, and the total uncertainty of each band and each elevation angle can be determined by the root mean square of the seven errors as

$$\begin{array}{}\text{(3)}& \mathit{ϵ}=\sqrt{{\mathit{ϵ}}_{\mathrm{fit}}^{\mathrm{2}}+{\mathit{ϵ}}_{\mathrm{temp}}^{\mathrm{2}}+{\mathit{ϵ}}_{\mathrm{topo}}^{\mathrm{2}}+{\mathit{ϵ}}_{\mathrm{SA}}^{\mathrm{2}}+{\mathit{ϵ}}_{\mathrm{SSA}}^{\mathrm{2}}+{\mathit{ϵ}}_{\mathrm{PF}}^{\mathrm{2}}+{\mathit{ϵ}}_{\text{2–4\hspace{0.17em}km}}^{\mathrm{2}}}.\end{array}$$

Figure 5Error budget of (a) UV and (b) VIS bands of the scanning cycle on 5 July 2015 at ∼16:26 UTC (SZA ∼64^∘, RAA ∼97^∘). The y axes refer to the relative error of O₄ DSCDs.

Download

Figure 6Same as Fig. 5, but for the scanning cycle on 7 December 2015 at ∼13:55 UTC (SZA ∼79^∘, RAA ∼39^∘).

Download

Examples of the error budgets of two measurement cycles for both wavelength bands are shown in Figs. 5 and 6. The cycle shown in Fig. 5 was measured in summer under relatively high aerosol load (AOD at 440 nm measured by the sun photometer around noon of that day was ∼0.2), while the cycle shown in Fig. 6 was measured in winter under relatively low aerosol load (AOD at 440 nm measured by the sun photometer around noon of that day was ∼0.015). In addition, the former cycle was measured under a smaller SZA compared to the latter one (64 and 79^∘, respectively), while the RAA was much larger than that of the latter (97 and 39^∘, respectively). The results show that contributions from different error sources are quite different in different measurement cycles, at different wavelengths, and at different elevation angles.

3.8 Inversion method

Aerosol extinction profiles are retrieved from the measured O₄ DSCDs of each scanning cycle. The measurements of the UV and VIS bands are retrieved separately. The measured O₄ DSCDs at the UV and VIS bands are fitted to the simulated O₄ DSCDs at 360 and 477 nm, respectively. In the retrieval, we assume the state of atmosphere to be stable during a scanning cycle and the distribution of aerosols to be homogeneous in horizontal direction. For a single scanning cycle, the measured O₄ DSCDs at the wavelength λ are denoted as a measurement vector

where M is the number of off-zenith measurements in each scanning cycle and is six in this study. ΔS_λ,1, ΔS_λ,2, …, ΔS_λ,6 are the O₄ DSCDs measured at O₄ wavelength band λ with the viewing elevation angles of 1, 2, 5, 10, 20, and 30^∘, respectively.

The simulated O₄ DSCDs corresponding to each possible aerosol extinction profile in X_LUT can be obtained from the lookup table. Similar to y_m, the simulation vector y_s for each possible profile x is denoted as

Aerosol extinction profiles can be derived by fitting the forward simulation to the measured O₄ DSCDs. Typically, the optimal solution can be determined by minimizing the cost function, which is defined as

where S_ϵ is the uncertainty covariance matrix. Assuming the measurements of each viewing elevation angle are independent, S_ϵ is a diagonal matrix and its diagonal elements are equal to the square of the total uncertainties of each elevation angle defined in Eq. (3),

Our cost function definition is similar to the cost functions used in many of the MAX-DOAS studies based on the OEM (e.g., Clémer et al., 2010; Frieß et al., 2016; Wang et al., 2016; Chan et al., 2017), but it only includes the item related to measurement error, while the item related to the a priori profile is omitted. This is because the a priori profile is not needed in our retrieval algorithm.

χ² indicates the difference between y_s and y_m. However, as the retrieval is ill-posed and the SNR of the measurements at the UFS is low, the single profile with the lowest χ² is not necessarily the one closest to the true profile. In order to overcome this limitation, we consider all the profiles in X_LUT with χ²(x) ≤ 1.5M (nine in this study) to be valid profiles and calculate the weighted mean profile as the optimal result. A profile with χ² ≤ M indicates that the measured and simulated O₄ DSCDs agree within the measurement errors, but in order to avoid underestimation of the measurement errors, we define the threshold as 1.5M. The weight of each valid profile for the calculation of the optimal solution is defined as

and the optimal solution can be calculated as

3.9 O₄ DSCD correction

Discrepancies between measured and simulated O₄ DSCDs are found in many other MAX-DOAS studies (Wagner et al., 2009, 2019; Clémer et al., 2010; Chan et al., 2015, 2017; Wang et al., 2016). The discrepancies are often explained by the systematic errors of the absorption cross section of O₄ as well as the radiative transfer simulation, and a correction is therefore necessary. Previous studies suggested multiplying a constant scaling factor (usually between 0.75 and 0.9) with the measured O₄ DSCD for all elevations to correct for the systematic error (e.g., Wagner et al., 2009; Clémer et al., 2010; Chan et al., 2015; Wang et al., 2016). Some recent studies suggested elevation-dependent scaling factors. Irie et al. (2015) suggested a set of scaling factors for 477 nm which gradually decreases with increasing elevation angle, varying from 0.984 for 1^∘ to 0.667 for 30^∘. Zhang et al. (2018) suggested a set of scaling factors for 360 nm, which also decreases with increasing elevation angle, varying from 1.02 for 1^∘ to 0.909 for 30^∘. Chan et al. (2017) derived a set of elevation-dependent scaling factors for 477 nm by comparing modeled and measured (relative) intensity, varying from 0.792 for 1^∘ to 0.957 for 30^∘. On the other hand, some other MAX-DOAS studies did not find it necessary to apply any correction to O₄ DSCDs. For example, Frieß et al. (2011) reported that for the MAX-DOAS measurements in an Arctic area, the measured and simulated O₄ DSCDs are in good agreement without any correction. Note that the scaling factor mentioned here refers to the ratio between simulated and measured O₄ DSCDs, which is opposite to some other studies.

In order to assess whether the O₄ DSCD correction is necessary for the MAX-DOAS measurements at the UFS, we compared the measured O₄ DSCDs to the simulated ones in the lookup table. Assuming our profile set (X_LUT) covers all possible aerosol profiles under cloud-free conditions, we derived the O₄ scaling factor for each elevation angle and each wavelength based on the statistical analysis. The AODs measured by the sun photometer were used to restrict the range of possible profiles.

Figure 7Distribution of simulated, measured, and corrected O₄ DSCDs in the (a) UV and (b) VIS bands of the scanning cycle on 7 December 2015 at ∼13:55 UTC (SZA ∼79^∘, RAA ∼39^∘). The x axes indicate the O₄ DSCDs measured (or simulated) at the elevation angle of 1^∘, while y axes represent the O₄ DSCDs measured (or simulated) at other elevation angles. Different colors indicate measurements at different elevation angles. The colored dots show the simulated O₄ DSCDs of all the possible profiles in the profile set (X_LUT). The data points of the profiles with AOD between 0 and 2 km (τ_2k(x)) vary between 50 % and 100 % of the total AOD measured by the sun photometer (τ_sp,λ) and are shown in bright colors, while the dots of the other profiles are shown in pale colors. The square markers represent measured O₄ DSCDs (corrected for the systematic errors caused by the topography simplification), and the error bars show the total uncertainties. The plus signs along the dashed lines show the measured O₄ DSCDs corrected with constant factors of 0.8, 0.9, 1.1, and 1.2. The triangle markers show the measured O₄ DSCDs corrected with the finally determined scaling factors listed in Table 6.

Download

Figure 8Cumulative distribution of the χ² of all the profiles in X_LUT for the scanning cycle on 7 December 2015 at ∼13:55 UTC (SZA ∼79^∘, RAA ∼39^∘). Dashed and solid curves refer to the results before and after the O₄ DSCD correction, respectively. Blue and red curves refer to the results of the UV and VIS bands, respectively. Note that the x axis is logarithmically scaled.

Download

Figure 7 shows the scattered plots of measured and simulated O₄ DSCDs of the scanning cycle on 7 December 2015 at ∼13:55 UTC. The measurements of both (a) UV and (b) VIS bands are shown. According to the cloud screening as well as the skycam images, this day was absolutely cloud free. Total AOD measured by the sun photometer at that time is 0.02 and 0.017 for the 360 and 477 nm bands, respectively. In each plot, the x axis indicates the O₄ DSCDs measured (or simulated) at the elevation angle of 1^∘, while the y axis represents the O₄ DSCDs at the other elevation angles. Different colors indicate measurements at different elevation angles. The simulated O₄ DSCDs (y_s(x)) of all the possible profiles in X_LUT are shown as colored dots. We assume the MAX-DOAS measurement of AOD between 0 and 2 km (denoted as τ_2k, ${\mathit{\tau }}_{\mathrm{2}\mathrm{k}}\left(\mathit{x}\right)=\mathrm{0.5}{\mathit{\sigma }}_{\mathrm{1}}\left(\mathit{x}\right)+\mathrm{0.5}{\mathit{\sigma }}_{\mathrm{2}}\left(\mathit{x}\right)+{\mathit{\sigma }}_{\mathrm{3}}\left(\mathit{x}\right)$) varies between 50 % and 100 % of the total AOD measured by the sun photometer (denoted as τ_sp,λ) in most cases, and the data points of the profiles fulfilling this assumption are highlighted. The measured O₄ DSCDs (already corrected for the systematic errors caused by the topography simplification) are plotted as square markers with error bars showing the total uncertainties. It is obvious that at most of the elevation angles, the measured O₄ DSCD does not agree with the simulations within the total error. As a result, at both the UV and VIS bands, no profiles in X_LUT satisfy the selection requirement (χ² ≤ 9; see dashed curves in Fig. 8). No profiles matching the measurement is unlikely to happen under such clear-sky conditions, hence implying a systematic error, and correction of the error is necessary.

In order to determine whether the O₄ scaling factor is constant for all elevations or dependent on the viewing elevation angles, we first assume it is constant and plot the corrected O₄ DSCD measurements in Fig. 7. The plus signs indicate the measured O₄ DSCDs corrected with constant scaling factors of 0.8, 0.9, 1.1, and 1.2. Furthermore, the corrected O₄ DSCDs should vary along the colored dashed lines if any other constant scaling factor is applied to the measurements. However, the forward simulation of O₄ DSCDs does not overlap with the dashed lines in most of the cases (especially for 5 and 10^∘ of the UV band), indicating that a constant O₄ scaling factor for all viewing elevation angles could not resolve the systematic error. Therefore, different scaling factors should be applied to different elevation angles.

In this study, the O₄ DSCD scaling factors for each viewing elevation angle and wavelength were determined through the statistical analysis of the long-term observations. We assume the scaling factor mainly depends on the viewing elevation angle, while being less sensitive to other factors such as solar geometry, aerosol load, temperature, etc.

Figure 7 shows that the simulated O₄ DSCDs at high elevation angles (e.g., 20 and 30^∘) vary in a very narrow range. Based on the assumption that X_LUT covers all possible aerosol profiles, the measured O₄ DSCDs should lie within the range. The scaling factor can be derived by taking the ratio of the simulated and measured values. As the simulated value varies in a narrow range, the uncertainty of the derived scaling factor should also be low. In order to have better statistics of the scaling factors, this method was applied to the long-term measurements. In addition, only the measurements taken under cloud-free and low-aerosol-load (${\mathit{\tau }}_{\mathrm{sp},\mathit{\lambda }}\le \mathrm{0.03}$) conditions were used, so as to avoid accounting for data contaminated by clouds in the analysis. Here it should be noted that measurements with AOD <0.03 are almost entirely found during winter due to the strong seasonal variation in aerosol load at the UFS. Subsequently, for the wavelength λ and the ith elevation angle of each scanning cycle, we calculate the variation range of the simulated O₄ DSCDs for all the profiles fulfilling $\mathrm{0.5}{\mathit{\tau }}_{\mathrm{sp},\mathit{\lambda }}\le {\mathit{\tau }}_{\mathrm{2}\mathrm{k}}\left(\mathit{x}\right)\le {\mathit{\tau }}_{\mathrm{sp},\mathit{\lambda }}$, which can be described as a set,

Only if $\max \left(Y_{\mathit{\lambda },i}^{\ast }\right)\le \mathrm{1.1}\times \min \left(Y_{\mathit{\lambda },i}^{\ast }\right)$, was the scanning cycle then taken into account. In most cases, measured O₄ DSCDs at high elevation angles are lower than simulated ones. Therefore we calculate the scaling factor from the minimum value in $Y_{\mathit{\lambda },i}^{\ast }$ to avoid overestimation of the scaling factor. The scaling factor derived from this scanning cycle is denoted as

where ΔS_λ,i is the measured O₄ DSCD (already corrected for the systematic errors caused by the topography). For the elevation angles of 5, 10, 20, and 30^∘ at the UV band and 10, 20, and 30^∘ at the VIS band, numerous scanning cycles from the long-term measurements fulfill the selection criterion, and hence there are sufficient samples of ${\mathit{\gamma }}_{\mathit{\lambda },i}^{\ast }$ for statistical analysis. We analyzed the frequency distribution of ${\mathit{\gamma }}_{\mathit{\lambda },i}^{\ast }$ of each elevation and each wavelength band. The result shows that the distributions of ${\mathit{\gamma }}_{\mathit{\lambda },i}^{\ast }$ follow the normal distribution function with small standard deviation. For instance, for the elevation angle of 20^∘, the standard deviations of UV and VIS bands are both ∼0.16. Subsequently, ${\mathit{\gamma }}_{\mathit{\lambda },i}^{\ast }$ with the maximum frequency was derived by Gaussian fit. The peak value was used as the scaling factor, which is denoted as ${\widehat{\mathit{\gamma }}}_{\mathit{\lambda },i}$.

For the low elevation angles (1 and 2^∘ at the UV band and 1, 2, and 5^∘ at the VIS band), as O₄ DSCD varies in a wide range, it is impossible to determine the scaling factor with the method mentioned above. However, it is found that in many scanning cycles, within the possible profiles in X_LUT, the simulated O₄ DSCDs at low elevation angles are well-correlated to those at the neighboring elevation angle. Therefore, once the scaling factor of the higher elevation angle is determined, we can derive an expected value of the O₄ DSCD at the lower elevation angle from the corrected O₄ DSCD at the higher one, and the scaling factor can be derived by taking the ratio of the expected value and the measured value.

For the wavelength λ and for each scanning cycle, a subset of X_LUT is defined as

and the elements of X^† are denoted as ${\mathit{x}}_j^{†}$. The corresponding simulated O₄ DSCD at the ith elevation angle is denoted as

A third-order polynomial regression is applied between $\mathrm{\Delta }{S}_{i,j}^{†}$ and $\mathrm{\Delta }{S}_{i+\mathrm{1},j}^{†}$. The regression function is denoted as g. Only if the correlation coefficient R² ≥ 0.98 is this scanning cycle taken into account. As the scaling factor of the (i+1)th elevation (${\widehat{\mathit{\gamma }}}_{\mathit{\lambda },i+\mathrm{1}}$) is already determined, the expected value of the O₄ DSCD at the ith elevation angle can be derived with the regression function

and the scaling factor derived from this scanning cycle is

Similar to the high elevation angles, the frequency distribution of ${\mathit{\gamma }}_{\mathit{\lambda },i}^{†}$ from all the available samples was analyzed by fitting to a Gaussian function. The peak value of ${\mathit{\gamma }}_{\mathit{\lambda },i}^{†}$ is used as ${\widehat{\mathit{\gamma }}}_{\mathit{\lambda },i}$. The scaling factor of the (i−1)th elevation is then derived in the same way. The scaling factors of 1 and 2^∘ at the UV band and 1, 2, and 5^∘ at the VIS band were determined using this method.

Table 6The finally determined O₄ DSCD scaling factors.

^* The O₄ DSCDs which are already corrected for the systematic errors caused by the topography simplification.

Download Print Version | Download XLSX

The determined scaling factors are listed in Table 6. The corrected O₄ DSCDs are indicated as triangles in Fig. 7. The result shows that except for the elevation angle of 1^∘, the simulated O₄ DSCDs are overestimated compared to the measured ones. It should be noted that the determination of the scaling factors is based on the measured O₄ DSCDs which are already corrected for the systematic errors caused by the topography simplification (discussed in Sect. 3.3). Comparing to the original measurements, the result still indicates that the simulated O₄ DSCDs at high elevation angles are overestimated. This result is opposite to the results of the other studies. At the moment we have no clear explanation for this finding, it might be related to the specific properties of the high-altitude station, e.g., the highly structured topography, horizontal gradients of the aerosol extinction, and systematic dependence of the surface albedo on altitude.

Figure 8 shows the cumulative distribution of χ² of all the profiles in X_LUT for the scanning cycle shown in Fig. 7. The distribution of χ² before and after the DSCD correction is shown as dashed and solid curves, respectively. The result indicates that for both UV (blue curves) and VIS (red curves) bands, the χ² values of most profiles in X_LUT are significantly lower after the correction. As a result, a number of profiles fulfill the selection criterion (χ² ≤ 9). Note that the AODs measured by the MAX-DOAS are still expected to be lower than the sun photometer observations due to the fact that the MAX-DOAS only reports the AOD below 2 km while the sun photometer covers the entire atmosphere.

4.1 Dependency of retrieval result on the threshold of cost function

As presented in Sect. 3.8, we consider all the profiles with χ²≤9 to be valid profiles, and the retrieved profile is defined as the weighted mean of all the possible profiles. In this section, we investigate the dependency of the retrieval result on the threshold of χ² by comparing the results calculated with different χ² thresholds. Taking the two measurement cycles mentioned in Figs. 5 and 6 for example, Fig. 9 (5 July 2015 at ∼16:26 UTC) and Fig. 10 (7 December 2015 at ∼13:55 UTC) show the weighted mean profiles, the variation range of valid profiles and the number of valid profiles corresponding to different χ² thresholds. The profiles are shown as colored curves which indicate the aerosol extinction coefficients in the three layers (i.e., σ₁, σ₂, and σ₃).

The results of both scanning cycles show that the retrieved profile is not sensitive to the threshold of χ² when there is a sufficient number of valid profiles (number of profiles exceeds ∼800 and ∼400 for UV and VIS, respectively; see the grey curves in Figs. 9 and 10). This is because profiles with larger χ² have lower weight (w). In addition, when the threshold value is increased, more profiles with both higher and lower aerosol extinction coefficients are taken into account. As a result, the variation range of valid profiles becomes larger but the weighted mean remains similar. The result shows that the retrieval with a χ² threshold of 9 is stable; therefore, it is used in the study.

Figure 11Weight distribution of valid profiles of the (a) UV and (b) VIS bands and results of the scanning cycle on 5 July 2015 at ∼16:26 UTC (SZA ∼64^∘, RAA ∼97^∘). The weight distributions of the aerosol extinction coefficients of the three layers (σ₁, σ₂, and σ₃) are shown as solid curves with different colors. The vertical dashed lines indicate the weighted mean aerosol extinction coefficient of the three layers (${\mathit{\sigma }}_{\mathrm{1}}\left(\widehat{\mathit{x}}\right)$, ${\mathit{\sigma }}_{\mathrm{2}}\left(\widehat{\mathit{x}}\right)$, and ${\mathit{\sigma }}_{\mathrm{3}}\left(\widehat{\mathit{x}}\right)$). The error bars indicate the weighted standard deviation calculated with Eqs. (16) and (17). The numbers on the error bars refer to the total weight (w) of the profiles covered by each error bar.

Download

Figure 12Same as Fig. 11, but for the scanning cycle on 7 December 2015 at ∼13:55 UTC (SZA ∼79^∘, RAA ∼39^∘).

Download

4.2 Estimation of the uncertainties of retrieved profiles

Still taking the two measurement cycles mentioned in Sect. 4.1 as examples, we analyzed the weight distribution of valid profiles; see Figs. 11 and 12. The distributions of aerosol extinction coefficients in the three layers (σ₁, σ₂, and σ₃) are shown as solid curves. For each layer, aerosol extinction coefficients of all the valid profiles are grouped, and the y axis refers to the total weight of each group. The three vertical dashed lines indicate the weighted mean aerosol extinction coefficient of each layer (i.e., σ₁, σ₂, and σ₃ of $\widehat{\mathit{x}}$). The result shows that the distributions of σ₁, σ₂, and σ₃ are all asymmetric for both the UV and VIS bands. In particular for the layer of 1–2 km (σ₃) at the UV band, the weight decreases monotonically with increasing aerosol extinction in both of the two cycles. Taking the cycle shown in Fig. 12 (7 December 2015 at ∼13:55 UTC) as an example, there are altogether 205 (12.8 %) and 120 (12.6 %) valid profiles with σ₃=0 in the UV and VIS bands, respectively. These profiles contribute total weights of 0.122 and 0.101 for the UV and VIS retrievals, respectively.

In order to estimate the uncertainty of $\widehat{\mathit{x}}$, we calculate the weighted standard deviations of σ₁, σ₂, and σ₃ of all the valid profiles. Due to the asymmetric distribution, the weighted standard deviations are calculated separately for the left (negative) and right (positive) sides. For the lth ($l=\mathrm{1},\mathrm{2},\mathrm{3}$) layer, denote the aerosol extinction coefficient of each profile as σ_l(x), then the weighted standard deviation of the left side is calculated from all the valid profiles with ${\mathit{\sigma }}_l\left(\mathit{x}\right)<{\mathit{\sigma }}_l\left(\widehat{\mathit{x}}\right)$,

and the weighted standard deviation of the right side is calculated from all the valid profiles with ${\mathit{\sigma }}_l\left(\mathit{x}\right)>{\mathit{\sigma }}_l\left(\widehat{\mathit{x}}\right)$,

The uncertainties of $\widehat{\mathit{x}}$ are indicated as error bars in Figs. 11 and 12. For each layer, the total weight of the profiles covered by the error bar is labeled in the charts. At the UV band, the total weight of the valid profiles covered by the uncertainties is 59 %–66 %, which is close to the standard normal distribution. However, the percentage can be up to 90 % at the VIS band. This is because the SNR of the measurement at the VIS band is higher. Therefore the retrieval of the VIS band has higher selectivity, and the weight is more concentrated to the mean value.

Figure 13Retrieval results of three synthetic profiles. The grey curves show the true profiles, with which the synthetic O₄ DSCDs were simulated. The blue and red curves represent the profiles retrieved using our LUT (lookup table) algorithm and a typical OEM (optimal estimation method) algorithm, respectively. The sold blue and red curves represent the profiles retrieved from the original synthetic data, and the dashed curves represent the profiles retrieved from the synthetic data with random noised added. The error bars of the blue curves indicate the uncertainties calculated by Eqs. (16) and (17). The dotted orange curves are the a priori profile used in the OEM retrieval.

Download

4.3 Retrieval of synthetic measurement data

In order to test the effectiveness of our retrieval algorithm, we generated some synthetic measurement data for the application to our algorithm. Figure 13 shows the results of three representative synthetic profiles at 360 and 477 nm. In each chart, the true profile is shown as the grey curve. Profile 1 is a tangent curve with aerosols distributed between 0 and 6 km above the instrument. The aerosol extinction decreases with increasing altitude, which is 0.04 km⁻¹ at surface level, ∼89 % at 2 km, and 50 % at 3 km. The total AOD is 0.12, of which ∼92 % is contributed from the altitude below 3 km. Profile 2 has a similar shape as Profile 1, but the aerosol extinction between 0.5 and 1 km above the instrument was enhanced. The aerosol extinction peaks at 0.75 km, and the average aerosol extinction coefficient between 0.5 and 1 km is larger than the bottom layer by ∼10 %. In addition, the aerosol extinction coefficients at other altitudes are increased by a factor of 2 compared to Profile 1. Profile 3 is an exponential profile. The total AOD is 0.12, the scaling height is 1.5 km, and the surface aerosol extinction coefficient is 0.03 km⁻¹.

We first simulated O₄ DSCDs at 360 and 477 nm with each profile. The solar position was set as SZA =60^∘ and RAA =60^∘, and the other parameters followed the settings used in calculating the lookup table listed in Table 5 (excluding the aerosol extinction coefficients above 2 km). In order to test the stability of the retrieval, we also generated a set of noisy data for each profile and each wavelength by adding random noise to the simulated O₄ DSCDs. We assume the measurement noise at all elevation angles is the same and follows a normal distribution with a standard deviation of 2 % of the DSCD of the lowest elevation angle. This noise level is realistic for the measurements at the UFS.

Aerosol profiles were then retrieved from both the original and noisy synthetic data using our algorithm. In the error estimation, the DOAS fitting error (ϵ_fit) was defined as the average values of the UFS measurements, while the other six kinds of errors followed the common settings presented in Sect. 3.7. O₄ DSCD correction was not applied. The solid and dashed blue curves in Fig. 13 show the profiles retrieved from the original and noisy data, respectively, and the error bars indicate the uncertainties calculated by Eqs. (16) and (17). The results show that for Profile 1 and Profile 3 our retrieval algorithm can reproduce the true profiles well from not only the original data but also the noisy data. For Profile 2, the retrieved profile cannot reproduce the elevated layer, but the error bar covers the aerosol extinction of the true profile. This is because the retrieval is ill-posed, which means the limited input information does not correspond to a unique profile with an elevated layer. Instead, many other profiles without the elevated layer can also fit the input information. Adding noise to the synthetic data can affect the retrieved aerosol extinction coefficients. However the influence is small in most cases. In addition, the noise can amplify the uncertainty of the retrieved profile. The results indicate that our LUT-based retrieval is stable.

We also retrieved the synthetic data using the bePRO profiling tool developed by BIRA-IASB (Clémer et al., 2010; Hendrick et al., 2014). It is an OEM-based algorithm and uses LIDORT as the forward model. In the retrieval of all six cases, the a priori profile was defined as an exponential profile with AOD =0.12 and scaling height =1.5 km, shown as the dotted orange curve in each panel of Fig. 13. The vertical grid was defined as 20 layers of 200 m thickness each. For Profile 1 and Profile 2, the uncertainty covariance matrix of the a priori profile (S_a) was defined as in Clémer et al. (2010) and Wang et al. (2014a): the diagonal elements corresponding to the bottom layer, S_a (1, 1), were set as the square of a scaling factor β (β=0.2) times the maximum partial AOD of the profiles; the other diagonal elements decrease linearly with increasing altitude to 0.2×S_a (1, 1); the off-diagonal elements of S_a were defined using Gaussian functions with correlation length γ=0.05 km. For Profile 3, as the difference between the true and a priori profiles is quite large, we set β=0.4 and γ=0.1 km, so that the constraint from the a priori profile is weaker. The measurement uncertainty covariance matrix (S_ϵ) was also defined as in most of the other MAX-DOAS studies so that S_ϵ is a diagonal matrix with variances equal to the square of the DOAS fitting error (${\mathit{ϵ}}_{\mathrm{fit}}^{\mathrm{2}}$). We defined ϵ_fit the same as in the LUT retrieval, but the other six error sources were not included. The retrieval parameters related to the radiative transfer simulation followed the settings of our LUT-based retrieval.

The results retrieved from the data with and without noise are shown in Fig. 13 as solid and dashed red curves, respectively. In all 12 retrieval cases, the O₄ DSCDs simulated with retrieved profiles are well-correlated to the input values (the relative root-mean-square error varies between 0.7 % and 4.7 %). However, as the retrieval is ill-posed, the retrieved profiles cannot reproduce the true profile well. Especially at high altitudes (above 1 km), the retrieved profiles are mostly dominated by the a priori profile. The OEM retrieval is also sensitive to measurement noise, which can be seen from the large variations in profile shape and aerosol extinction. The results indicate that the LUT-based algorithm is much more suitable for measurements with low SNR.

Figure 14Comparison of AODs at (a, b) 360 nm and (c, d) 477 nm measured by the MAX-DOAS and sun photometer. The charts on the left side (a, c) show the daily and monthly averaged time series, whereas the scatter plots on the right side (b, d) show the hourly averaged results. The AOD measured by MAX-DOAS refers to the vertical range between 0 and 2 km (i.e., ${\mathit{\tau }}_{\mathrm{2}\mathrm{k}}\left(\widehat{\mathit{x}}\right)=\mathrm{0.5}{\mathit{\sigma }}_{\mathrm{1}}\left(\widehat{\mathit{x}}\right)+\mathrm{0.5}{\mathit{\sigma }}_{\mathrm{2}}\left(\widehat{\mathit{x}}\right)+{\mathit{\sigma }}_{\mathrm{3}}\left(\widehat{\mathit{x}}\right)$) above the instrument (i.e., 2650–4650 m a.s.l.). The measurements were available during daytime with SZA <85^∘ and cloud-free conditions. The AOD measured by the sun photometer refers to the total AOD, and only the measurements during 10:00–14:00 UTC were used due to their accuracy. The daily and monthly averaged results are calculated from all the available hourly averaged AODs; therefore they are not real monthly and daily averages. The error bars of the MAX-DOAS data refer to the averages of the uncertainties calculated by Eqs. (16) and (17). A few data points are outside the scatter plots.

Download

4.4 Comparison to sun photometer measurements

Figure 14 shows the comparison of AODs measured by the MAX-DOAS and sun photometer during the entire study period. The seasonally averaged AODs measured by both instruments are listed in Table 7. As the AOD measured by the MAX-DOAS refers to the AOD between 0 and 2 km while the AOD measured by sun photometer refers to the total AOD, the sun photometer results should be larger. Despite the difference, the time series (panels (a) and (c) of Fig. 14) show that the AODs measured by both instruments have a similar seasonal variation with higher AOD in summer and lower AOD in winter. The monthly average data show that the difference between the AODs measured by the MAX-DOAS and sun photometer is much larger in summer, which coincides with the ceilometer profiles shown in Fig. 1 which indicate much higher aerosol extinction coefficients above 2 km in summer. The underestimation of the MAX-DOAS may also be related to the decreased sensitivity of measurement at higher altitudes.

Table 7Seasonally averaged AODs measured by the MAX-DOAS and sun photometer at the UFS. The AODs measured by the MAX-DOAS refer to the AODs between 0 and 2 km above the instrument (i.e., 2650–4650 m a.s.l.), and the measurements were available during the daytime with SZA <85^∘ and no cloud; the AODs measured by sun photometer refer to the total AOD, and the measurements were only available during 10:00–14:00 UTC. The results listed in the table are calculated from all the available hourly averaged AODs.

Download Print Version | Download XLSX

The correlation between hourly averaged AODs measured by the MAX-DOAS and sun photometer is shown in Fig. 14b, d. AODs show a general agreement at the UV and the VIS bands with correlation coefficients of R=0.733 and 0.798, respectively. However, AODs from the MAX-DOAS are lower; consequently the slope of the regression lines is 0.5308 and 0.3556 for the UV and VIS bands, respectively. As the MAX-DOAS only reports AODs below 2 km while the sun photometer measures the total AODs, the MAX-DOAS AODs are indeed expected to be lower. This is in particular true in cases of large AODs due to very strong convection of polluted air masses from the valley and/or the presence of Saharan dust layers. Then, particles are often transported beyond the range of the MAX-DOAS measurements and the disagreement is largest. This feature might be strengthened by the decreased sensitivity of the MAX-DOAS measurement at higher altitudes, so that the upper part of an aerosol layer is missed. In addition, a few data points lie above the 1:1 reference lines. This might be explained by the inhomogeneous distribution of aerosols in the horizontal direction. The light paths of the MAX-DOAS and the sun photometer are different. The MAX-DOAS measures scattered sunlight while the sun photometer derives the AOD from direct sun measurements. Therefore, when the aerosol load along the light path of the MAX-DOAS is higher than that of the direct sun measurement, the AOD measured by the MAX-DOAS may exceed the AOD measured by the sun photometer. For most of these points, the difference between the results of the two instruments is within their uncertainty ranges; i.e., the disagreement is probably due to the measurement and retrieval errors.

Figure 15Seasonal average aerosol extinction profiles for (a) 360 and (b) 477 nm derived from the long-term measurement results. The error bars represent the averages of the uncertainties calculated by Eqs. (16) and (17).

Download

4.5 Temporal variation in aerosol characteristics

The seasonally averaged aerosol extinction profiles derived from the long-term measurements are shown in Fig. 15. The result indicates that the aerosol load is high in summer and low in winter, which coincides with the ceilometer results shown in Fig. 1. The seasonal pattern can be explained by the higher biogenic emissions from vegetation in summer. Moreover, the mixing layer is higher in summer; thus anthropogenic aerosols are more likely dispersed to upper altitudes. The shape of the profiles also agrees with the ceilometer results that the averaged aerosol extinction decreases with increasing altitude in all seasons – taking into account the coarse vertical resolution of the MAX-DOAS. In addition, Fig. 15 shows a much larger vertical gradient at 360 nm in summer. This might be explained by the lower sensitivity of the UV measurement for higher altitudes due to the more decreased visibility at shorter wavelengths.

Figure 16Comparison of seasonal average aerosol extinction coefficients at 360 and 477 nm in the bottom layer (0–0.5 km above the instrument, σ₁). The colored bars show the average aerosol extinction coefficients of the four seasons (equal to the bottom values shown in Fig. 15). The grey square markers indicate the ratios between the aerosol extinction coefficients at 360 and 477 nm.

Download

We compared the seasonally averaged aerosol extinction coefficients at 360 and 477 nm in the bottom layer (0–0.5 km above the instrument, σ₁); see Fig. 16. The averaged aerosol extinction coefficients are shown as bar charts. The ratio between the aerosol extinction coefficients at 360  and 477 nm is indicated by the grey curve. The result shows that the aerosol extinction coefficient ratio between 360  and 477 nm is significantly higher in summer than in the other seasons.

From these ratios the Ångström exponent (AE) can be calculated using the seasonally averaged surface aerosol extinction coefficients at 360 and 477 nm. The results are listed in Table 8. The seasonal averaged AEs of 380–500 nm from the AERONET measurements at Hohenpeißenberg from April 2013 to February 2016 are also listed for comparison. The result shows that both the UFS and Hohenpeißenberg measured the highest AE in summer and the lowest in winter. The AE at the UFS is in general lower than that measured at Hohenpeißenberg with a smaller difference in summer. This can be explained by the different altitude of the two sites. As the AERONET station at Hohenpeißenberg is located at ∼950 m a.s.l., a larger contribution of anthropogenic aerosols is expected. The extremely low AE at the UFS in spring, autumn, and winter agrees with the result measured at a plateau site (Lhasa, China, 3688 m a.s.l.) reported by Xin et al. (2007). The annual mean AE at that site is reported to be 0.06±0.31, which is significantly lower than that measured at low-altitude sites, especially urban and forest sites. In general, a smaller AE implies larger aerosol particle sizes (Dubovik et al., 2002). The increased AE at UFS in summer indicates a larger contribution of fine particles. The result is consistent with the fact that the particle size of biogenic secondary aerosols is in general smaller than ice particles transported from the lower altitudes to upper altitudes in summer.

Table 8Seasonal average Ångström exponents (AEs) obtained from MAX-DOAS near-surface measurements (0–0.5 km above instrument) and from AERONET measurements at Hohenpeißenberg. The results of the MAX-DOAS are calculated from the ratios between the seasonal average aerosol extinction coefficients at 360 and 477 nm (i.e., the ratios shown in Fig. 16). The results of AERONET are the seasonal average values of AEs (380–500 nm) at Hohenpeißenberg from April 2013 to February 2016.

Download Print Version | Download XLSX