Improved retrieval of SO 2 from Ozone Monitoring Instrument : residual analysis and data noise correction

Improved retrieval of SO2 from Ozone Monitoring Instrument: residual analysis and data noise correction H. Yan, L. Chen, J. Tao, L. Su, and D. Han State Key Laboratory of Remote Sensing Science, Jointly Sponsored by the Institute of Remote Sensing Applications of Chinese Academy of Sciences and Beijing Normal University, Beijing, China Graduate University of the Chinese Academy of Sciences, Beijing, China


Introduction
Sulfur dioxide (SO 2 ) is an important index to urban atmospheric quality evaluation.
The spatiotemporal changes of SO 2 have a significant impact on the environment and the climate.The natural sources of SO 2 are sulfur minerals decomposition, oxidation of dimethyl sulfate from marine phytoplankton, and degassing and eruptions of volcanoes.Anthropogenic activities also increase the SO 2 concentration in the atmosphere.These anthropogenic activities include the smelting of sulfur ore, combustion of coal, emission of motor vehicle, and industrial activities (Cullis and Hirschler, 1980).
Since the Nimbus-7 Total Ozone Mapping Spectrometer's (TOMS) first sighting of SO 2 cloud from the El Chich ón's eruption in 1982 (Krueger, 1983), satellite remote sensing technology has begun to play an important role in the field of atmospheric monitoring.Later, Global Ozone Monitoring Experiment (GOME), SCanning Imaging Introduction

Conclusions References
Tables Figures

Back Close
Full Absorption spectroMeter for Atmospheric CartograpHY (SCIAMACHY), and Ozone Monitoring Instrument (OMI) were developed, and all these instruments have high SO 2 monitoring capability.Particularly, OMI, which was launched on the EOS/Aura platform in July 2004, combines the hyperspectral measurements of GOME and SCIAMACHY with improved nadir spatial resolution (13 × 24 km 2 ), and realizes the daily global monitoring of short-lifetime SO 2 .It measures solar radiation backscattered by the earth's atmosphere in the sun-synchronous polar orbit with 1:45 p.m. local equator crossing time, and makes simultaneous measurements in a swath of width about 2600 km, divided into 60 pixels (Levelt et al., 2006).OMI has high sensitivity to anthropogenic SO 2 , enabling the acquisition of information of near-surface SO 2 emission; it plays an important role in near-surface anthropogenic SO 2 emission monitoring (Krueger et al., 2002;Krotkov et al., 2008;Carn et al., 2007;Li et al., 2010;Yan et al., 2012).Methods for retrieval of SO 2 from satellite instruments have been developed and applied to detect volcanic SO 2 for aviation (Yang et al., 2007;Krotkov et al., 2006;Rix et al., 2010;Lee et al., 2008;Corradini et al., 2009;Realmuto et al., 1994).Among the ultraviolet inversion algorithms of SO 2 , Differential Optical Absorption Spectroscopy (DOAS) (Platt, 1994;Platt and Stutz, 2008) and Linear Fit algorithm (LF) (Yang et al., 2007) are the ones sensitive to large volcanic SO 2 in the upper troposphere or low stratosphere (Yang et al., 2007;Rix et al., 2010;Lee et al., 2008).Band Residual Difference algorithm (BRD) provides unprecedented measurement sensitivity to nearsurface small SO 2 emission (Krotkov et al., 2008;Carn et al., 2007;Li et al., 2010;Yan et al., 2012).The BRD algorithm selects four wavelengths between UV 310.8 nm and 314.4 nm (310.8 nm, 311.9 nm, 313.2 nm, and 314.4 nm), which include the wave crests and wave troughs of SO 2 absorption (Fig. 1).It constitutes three wavelength pairs (P1 = 310.8-311.9nm, P2 = 311.9-313.2nm, and P3 = 313.2-314.4nm), and uses OMI O 3 column amount and Lambertian Effective Reflectivity (LER) to estimate the total vertical column amount SO 2 (Krotkov et al., 2006).With high sensitivity to SO 2 , BRD can detect the changes of small anthropogenic SO 2 emission in the Planetary Boundary Layer (PBL), which has a significant impact on environment monitoring Introduction

Conclusions References
Tables Figures

Back Close
Full  We chose China as our study area and compared the results between the optimized results and the OMI SO 2 PBL products.Results show that former has higher precision and reasonable SO 2 vertical total amount distribution.For the present study, we used data from the National Aeronautics and Space Administration (NASA) Goddard Earth Sciences Data and Information Services Center (GESDISC).

Data noise sources and residual analysis
By using the above-mentioned BRD SO 2 retrieval method, OMI level 1 solar irradiance and earth radiance data (Van den Oord et al., 2002), and OMI level 2 O 3 column amount data (Bhartia and Wellemeyer, 2002), we obtain the vertical total column SO 2 in the atmosphere (DU = Dobson Units or milli atm cm, 1 DU = 2.69 × 10 16 molecules cm −2 ). Figure 2 shows the SO 2 column amount results by using the OMI level 1 solar irradiance (orbit 18 570), earth radiance data (orbits 18 565 and 18 566), and OMI level 2 O 3 products (orbits 18 565 and 18 566) on 11 January 2008.Without solar irradiance correction and residual correction, the retrieved SO 2 column amount results show several clear stripes, which are outlier values in the atmosphere SO 2 column data as a function of the viewing angle (Fig. 2a-d).These regularly Introduction

Conclusions References
Tables Figures

Back Close
Full distributed stripes noises mask the effective signals, including SO 2 absorption information.Therefore, to get the reasonable global distribution of SO 2 and enable OMI SO 2 application, it is necessary to mitigate these stripes noise.OMI uses the sun-synchronous polar orbit with a daily global coverage and a solar irradiance corresponding to about 14 earth radiances in one day measurement.When the same solar irradiance data (orbit 18 570) are applied to different earth radiance data (orbits 18 565 and 18 566) in one day, the SO 2 column amount distribution shows a clear along-track stripes error because of the invalid value in the solar irradiance data.We used the nearest neighbor interpolation method to remove the invalid value in the solar irradiance data, and the clear stripes error can be removed, as shown in Fig. 2c, d.However, even after such reprocessing, the SO 2 column amount remains to have along-track noise.Different earth measurement orbits in one day show the coincident stripes distribution, similarly performing at some certain observation angles, as shown in Fig. 2c, d.Considering the OMI observation mode, these SO 2 alongtrack stripes mainly come from solar irradiance noise, which is affected by dark signal, diffuser features, and signal noise, among others (Veihelmann and Kleipool, 2006).These noises in solar irradiance are not constant; rather, they change over time.
Apart from solar irradiance noises, the earth radiance data noises also introduce cross-track noise into the SO 2 column amount results.As OMI sensor ages, the earth radiance data are affected by the row anomaly effect since late 2008.This row anomaly effect causes four distinct errors on earth radiance: a multiplicative error by blockage effect, which causes the radiance value to decrease; a wavelength shift error; a stray earthlight related additive error, which causes the radiance value to increase; and a stray sunlight related additive error, which causes the radiance value to increase (http://www.knmi.nl/omi/research/product/rowanomaly-background.php).The OMI row anomaly is dynamic, which means that it varies with time.
In the process of SO 2 retrieval, the residual, as an intermediate value, is the result of the measured reflection subtracted from the simulated reflection at the top of the atmosphere, and determines the final SO 2 result.It is affected by many factors, including Introduction

Conclusions References
Tables Figures

Back Close
Full solar irradiance noise, earth radiance noise, O 3 data noise, SO 2 and O 3 absorption spectrum, satellite observation geometry, and other error sources.To mitigate these noise interferences, background correction for residual is needed before calculating the SO 2 vertical column amount.
The DOAS retrieval algorithm uses a reference sector method by subtracting the presumably SO 2 free reference data on the same day at the same latitude (Khokhar et al., 2005;Richter et al., 2006).However, the forward model errors or satellite measurement errors that affect the SO 2 retrieval results are dynamic over the world and correlate with time, observation geometry, O 3 amount, and surface reflectivity, among others.Therefore, using the fixed reference sector for the empirical correction of SO 2 retrieval possibly bring new error sources.Yang K developed a sliding median method for residual correction, which has been applied to OMI SO 2 products (Krotkov et al., 2008;Yang et al., 2007).It calculates the median residual for each cross-track pixel from a sliding group pixels covering about 30 • latitude along the orbit track, and then subtract the sliding median value for each spectral band and each cross-track position, to reduce the error interference for the centered pixels.This method can reduce any cross-track and along-track biases.This method enables the relative SO 2 distribution, and not the absolute SO 2 results.As seen in Fig. 2e, f, after residual correction, the along-track stripes were largely reduced.In addition, the quality of the OMI SO 2 PBL data on 11 January 2008, is relatively reliable and has low signal-to-noise ratio; however, since 2009, as the OMI sensor ages, the SO 2 products have been seriously affected by row anomaly.Moreover, data noises emerge largely in China, which cannot be effectively corrected by the current SO 2 BRD retrieval algorithm (Krotkov et al., 2006) (Fig. 3).These high-value noises were clearly erroneous, making the SO 2 results unreliable on that day.We provided a new scheme for the BRD SO 2 retrieval algorithm, which reduced the BRD SO 2 results noise, improved the precision of the BRD SO 2 results, and enabled the reasonable SO 2 vertical column amount distribution.
By selecting an area free of SO 2 , over the North Pacific Ocean area (0-30 • N, 140-Introduction

Conclusions References
Tables Figures

Back Close
Full Figure 4 shows that the median value of the residual for residual correction from a sliding group pixels covering 20 • or 10 • latitude is consistent with the results covering 30 • latitude in the cross-track direction on 16 January 2008; meanwhile, the median residual from the 5 • latitude sliding group pixels is not sufficient to include the background error.When the selected pixel is near the terminator with high solar zenith angle, the terminator data with high noise for residual correction will result in the worsening of the selected pixel retrieval result.Hence, it is important to determine the residual correction area.As shown in Fig. 3, with 30 • latitude residual correction area, the OMI SO 2 PBL product yields clear cross-track noises in January 2009.From the residual analysis on 12 January 2009, as shown in Fig. 5, the median value from 30 • latitude residual correction area has an abnormally low value in the 40 pixel position at the cross-track direction, which largely deviates from the uncorrected residual for the centered pixel.With such outlier median value for residual correction, the SO 2 retrieval results will show spike value at the cross-track direction.The median value from 20 • latitude residual correction area also has a low value in the 40 pixel position, whereas 10 • latitude median value did not record a low value.Therefore, the median value from 10 • latitude residual correction area has a relatively better residual correction effect.
In this study, we used the modified residual correction scheme to reduce the SO 2 column amount noise and improve the precision of the BRD SO 2 algorithm retrieval results.In detail, by using the TOMS forward model TOMRAD (Dave, 1964), we calculated the residual at four wavelengths between UV 310.8 nm and 314.4 nm (310.8 nm, 311.9 nm, 313.2 nm, and 314.4 nm) and assumed the air mass factor (AMF) value as a constant 0.36; we used the SO 2 absorption cross-section data at a constant temperature (273 K) (Bogumil et al., 2003) and calculated the median residual value from a sliding group of 10 • latitude residual correction area.Comparing with the mean residual value, the median value is less sensitive to the outlier value.When the selected pixels are near the terminator, we decrease the residual correction area.Bad pixels determined by residuals are filtered before median residual value calculation.Finally, Introduction

Conclusions References
Tables Figures

Back Close
Full by subtracting the corresponding median residual value, we obtained the corrected residual and then used these corrected residual to retrieve the SO 2 column amount.

Results and discussion
By selecting OMI level 1 and level 2 products in China on 9, 11, 12 and 16 January 2008, we analyzed the space distribution coherence and correlation of the SO 2 column amount between the modified algorithm results and the OMI released SO 2 PBL products.The period chosen is during the Chinese winter, which has a low near-surface temperature, with the SO 2 average conversion rate in the atmosphere being lower than that in the summer (Eatough et al., 1982;Khoder, 2002); during this period, anthropogenic activities are increasing (like coal heating) -factors that make SO 2 emission higher than in the summer -which is a more effective reflection of regional anthropogenic SO 2 emission and is easily detected by satellite.In this study, we chose China as the research area.Since the last century of the 1980s, after China's reforms and opening up, the Chinese economy has rapidly boomed and changed from an agriculture country to an industrial one.However, industrialization and urbanization caused the large emission of pollution gases, greenhouse gases, and particulate matter; the continuous degradation of air quality; and even caused serious environmental pollution incident such as acid rain, haze, and photochemical smog (Chen et al., 2009).As shown in Fig. 6, in the cross-track direction, the modified BRD SO 2 algorithm results are consistent with the OMI SO 2 PBL product, having similar crest and trough variation in January 2008.However, at the crest of the SO 2 vertical column amount, the absolute value of the modified BRD SO 2 algorithm results is slightly lower than the OMI SO 2 PBL product.This can be caused by several factors, such as background biases subtracted larger, zenith angle interval difference in the forward model, and SO 2 absorption convolving with instrument response function, among others.
We filled the China SO 2 column amount space distribution based on 0.125 • × 0.125 • latitude-longitude grid, and analyzed the SO 2 column amount space distribution Introduction

Conclusions References
Tables Figures

Back Close
Full coherence and correlation between the modified algorithm and the OMI level 2 SO 2 PBL product.As shown in Fig. 7, OMI SO 2 PBL product data have high signal-to-noise ratio in January 2008, and the spatial distribution between the modified algorithm and the OMI level 2 SO 2 PBL product has better coherence in the high-value sector than in the low-value sector.The high-value sector is mostly concentrated in Eastern China (e.g., Shandong, Hebei, Henan, and Tianjin) and in the southwest areas of China (e.g., Chongqing and Chengdu).The low-value areas of the modified algorithm are mainly distributed in China Tibet, while OMI level 2 SO 2 PBL products do not show clear low value in the same area.With mass data correlation analysis, the SO 2 column amount of the modified algorithm and the OMI level 2 SO 2 PBL product has good coherence in January 2008 (correlation coefficient R ranges from 0.6674 to 0.8843).However, since 2009, with the aging of the OMI, satellite signal has begun to attenuate and data noise largely increased; OMI level 2 SO 2 PBL products were affected by the row anomaly effect (Fig. 3).In Fig. 3, OMI level 2 SO 2 PBL products appear to have abnormally high noise in China, which causes the erroneous SO 2 spatial distribution over Inner Mongolia, Shaanxi, Chongqing, Guiyang, and Guangxi.This is unfavorable for OMI SO 2 PBL product application.By using the modified algorithm proposed above, we improved the result of the SO 2 column amount.The modified algorithm effectively removed the unusually high-value noise and obtained the reasonable SO 2 spatial distribution result (Fig. 8).In areas without abnormally high noise, the modified algorithm results and the OMI SO 2 PBL products are consistent, both reflecting the high SO 2 emission of the Chengdu area.This proves that the modified algorithm not only reflects the relativity distribution of SO 2 , but also yields higher precision.
By selecting the North Pacific Ocean area (15-20 • N, 135-150 • E) where we presumed that there are SO 2 -free data, we compared the SO 2 results precision between the modified algorithm and the OMI level 2 SO 2 PBL product (Table 1).Without manmade SO 2 emission, the SO 2 column amount should be near zero in this area (Chin et al., 2000).However, because of the sensor measurement and retrieval algorithm errors, the SO 2 column amount value suffered varying degrees of noise interference

Conclusions References
Tables Figures

Back Close
Full over the selected area.As shown in Table 1, the precision (standard deviation, σ) of the OMI level 2 SO 2 PBL product data in the selected area is about 1.3 DU, increasing to about 3.4 DU in January 2009, which made the SO 2 results unreliable.After the modification of the algorithm, the SO 2 column amount precision was improved to about 1 DU over the same area during the same period.Meanwhile, temporal and spatial averaging can reduce noise (Krotkov et al., 2008).

Conclusions
In this study, we analyzed the OMI SO 2 main noise sources, solar irradiance, and earth radiance, which are regarded as the main measurement error sources.Since 2009, the increasing row anomaly effect has been unfavorable for OMI SO 2 PBL product application.Based on the current BRD SO 2 algorithm and by analyzing the different residual correction effects and determining the best residual adjustment area, we effectively reduced the data noise and realized the optimization of the BRD algorithm.In the present study, we selected China as our study area and compared the results between the optimized and the OMI SO 2 PBL products.Results show that they are consistent with each other in January 2008; however, in January 2009, the modified algorithm results have shown higher precision and more reliable spatial distribution.Satellite remote sensing technology has incomparable advantages such as a wider coverage area, shorter revisit period, and spatial continuity.However, satellite observation with instantaneous field of view can only obtain transient overpass information; it cannot reflect the continuous temporal variation of atmospheric conditions.Therefore, even with daily global coverage, transient OMI satellite observation cannot precisely reflect the variation of atmospheric conditions in one day.With a short lifetime in the atmosphere, the trace gas SO 2 can rapidly transform to sulfate in several days or even in a few hours (Eatough et al., 1982;McGonigle et al., 2004;Khoder, 2002), demonstrating its clear regional effect.Additionally, SO 2 concentration in the atmospherewhich is affected by natural and anthropogenic emission -varies largely in one day.

Conclusions References
Tables Figures

Back Close
Full Therefore, it is necessary to develop joint retrieval through multi-sensor fusion, or combine with atmospheric transport and chemistry model, to obtain the continuous variation of SO 2 in the atmosphere.
In addition to the measurement error, the retrieval algorithm error, which was not discussed in this study, can also bring about more complex noises to the SO 2 result.In this study, we assumed that surface reflectivity is constant for all the retrieval wavelengths.However, such assumption is only a hypothesis for retrieval simplification since, in reality, surface reflectivity varies with wavelength.The gas absorption cross-section data we used were at a constant temperature (273 K), thus temperature variation effect was not considered.In the forward model, there is an overly simplified treatment of many complex processes occurring in a real atmosphere, including Mie scattering by clouds and aerosols.Moreover, the nonlinear effect of the algorithm itself can result in SO 2 result underestimation.All these errors can altogether produce high noise that masks the real SO 2 distribution.
Therefore, although satellite technology can obtain pollutants spatial distribution, source and sink distribution, and transport route, the complex atmosphere conditions cannot be precisely simulated by a simple mathematics, physics model.The quantitative retrieval of satellite monitoring, especially near-surface SO 2 concentration retrieval, needs further study in the future.Introduction

Conclusions References
Tables Figures

Back Close
Full Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | slant column less than 2 DU; we analyzed the different residual correction area effects.
Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Screen / Esc Printer-friendly Version Interactive Discussion Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Fig. 2. SO 2 stripes caused by solar irradiance and SO 2 results with residual correction.

Improved retrieval of SO 2 from Ozone Monitoring Instrument: residual analysis and data noise correction H. Yan
1,2 , L. Chen 1 , J. Tao 1 , L. Su 1 , and D. Han 1 1 State Key Laboratory of Remote Sensing Science, Jointly Sponsored by the Institute of Remote Sensing Applications of Chinese Academy of Sciences and Beijing Normal University, Beijing, China 2 Graduate University of the Chinese Academy of Sciences, Beijing, China

Table 1 .
Data precision of OMI level 2 SO 2 PBL column amount and modified algorithm SO 2 results in the North Pacific Ocean area.