Remote Sensing Data Combinations - Superior Global Maps for Aerosol Optical Depth

Stefan Kinne

MPI-Meteorology, Hamburg Abstract. More accurate and complete measurement data-sets are needed to constrain the freedom of global modeling and raise confidence in model predictions. Frequently, many measurement data of the same and a related property exist with often quite different limitations to regional or seasonal or vertical coverage and strength. In that case, there is a need to combine the strengths of individual data sources for superior products. As a demonstration, different aerosol optical depth (AOD) multi-annual data sets from remote sensing have been combined into AOD maps that score (in reference to quality data from sun-photometry) better than any individual satellite retrieval. Further improvements are achieved by adding statistics from sun-photometry to the satellite composite. The global average mid-visible AOD of this remote sensing composite is near 0.13 annually, with lower values during northern hemispheric fall and winter (0.12) and larger values during northern hemispheric spring and summer (0.14). This data-set also reveals characteristic deficiencies in global modeling. Modeling tends to overestimates AOD over the northern mid-latitudes and to underestimate AOD over tropical and sub-tropical land regions. Also noteworthy are AOD underestimates by modeling in remote oceanic regions, though only in relative sense as AOD values in that region as small. The AOD remote sensing data composite is far from perfect, but it demonstrates the extra value of data-combinations.

Introduction Aerosol remote sensing from space is predominantly based sensor-data of reflected sun-light in solar spectral regions, where the attenuation by trace-gases can be neglected. However, these aerosol retrievals are difficult for three major reasons.

1. cloud contamination: The solar reflection attributed to aerosol is small compared to that of clouds and identifying cloud free and cloud-influence free (e.g. cloud shadows) regions is a challenge, especially with sensor limitations to spatial resolution and non-nadir viewing angles (e.g. cloud shades).

2. surface contributions: The solar reflection attributed to aerosol can be smaller than surface signals. Thus, surface albedo (also as function of the sun-elevation) needs to be known at high accuracy. To minimize the surface albedo problem innovative methods are applied. They rely on spectral dependencies (Kaufman et al., 1997), multi-angular views (Martonchick et al., 1998), polarization (Deuze et al., 1998) or retrievals in the UV (Torres et al, 2002). Higher and variable surface albedos still remain the major reason that most aerosol satellite products have no or only limited coverage over land.

3. a-priori assumptions: The relationship that associates changes of solar reflection to aerosol amount at (cloud-free conditions) is modulated by aerosol composition and even atmospheric environment. With usually limited information from

1

sensor data, many a-priori assumptions are required, in particular on aerosol absorption and aerosol size. Some of these assumptions have been locally and/or seasonally validated, but their regional (or even global) and annual application is rarely justifiable. The availability of remote sensing data with respect to aerosol properties is uneven, Table 1 illustrates that the most commonly retrieved aerosol property is the (mid-visible) aerosol optical depth (AOD, representing aerosol amount) over oceans. Table 1. Aerosol properties, availability and associated techniques in satellite remote sensing

property ocean land availability technique availability technique amount Good solar reflection limited multi-directional absorption Poor glint poor UV-spectral Size Limited multi-spectral poor Polarization shape Poor polarization poor multi-directional altitude Poor multi-directional limited lidar

The different approaches listed in Table 1 can provide - next to AOD - important constrains on other aerosol and environmental properties. Thus, among the different available AOD maps from satellite remote sensing quality differences can be expected. However, since more capable sensors are not always matched with better retrieval assumptions, there is ambiguity which satellite AOD products to believe. Thus, there is a strong community interest to compare and assess available AOD data-sets and to provide needed recommendations for AOD measurement use (e.g. for model input or evaluation).

Satellite AOD data-sets

The use of different and often complementary techniques is desirable, but at the same time complicates data comparisons, as different sensors, retrievals methods and assumptions are applied. To minimize additional complications by differences in data-sampling, only monthly mean properties of multi-annual AOD data-sets are considered. The data-set comparison includes suggestions by MODIS collection 5 and 4 (2000-2005), MISR (2000-2005), TOMS (1979-2001) new and old processing, POLDER (1987, 2002) and AVHRR NOAA (1981-1990) and AVHRR GACP (1984-2001). Time periods with enhanced stratospheric aerosol loading (e.g. after the El Chichon 1982-1985 and after the Mt.Pinatubo 1991-1994 volcanic eruptions) are excluded from these averages. Annual global AOD maps of these eight remote sensing data products are compared to data from ground-based monitoring in Figure 1. Background information is provided in Table 2.

Despite some similarity in major AOD annual patterns, the differences among the available AOD data are large (Liu 2008). Already for annual AOD retrievals, locally the diversity is comparable and often larger than the AOD average. In need for a reference, local statistics from ground based monitoring is applied. Ground-based monitoring of direct solar attenuation (a transmission) by sun-photometry has the advantage that AOD can be directly measured without a need to prescribe aerosol composition and (unless aerosols are large) aerosol size. In addition, background contributions are well defined.

2

Figure 1. Annual global mid-visible AOD maps from remote sensing (at 0.55μm wavelength). Sun-photometer data (aer - enlarged for visual purposes) of ground-based monitoring networks by AERONET, SKYNET and GAW are compared to satellite sensor retrievals of MISR (MIS), of MODIS collections 5 (Mc5) and 4 (Mc4) - of both TERRA and AQUA platforms), of AVHRR by NOAA (AVn) and by GACP (AVg), of TOMS older (TOo) and newer (TOn) data processing and of POLDER (POL). Numbers at labels display annual averages of locations with (non-zero) data Table 2. Multi-annual available AOD data-sets from remote sensing, temporal coverage, literature references, major data limitations and recognized (pos + or neg -) biases

sensor year references limitation bias Aer AERONET 96-06 Holben 98 local, land only Mc4 MODIS,T+A 00-05 Tanré 97, Kaufman 97 no deserts + over land Mc5 MODIS,T+A 00-05 Tanré 97, Remer 05 no deserts MIS MISR 00-05 Kahn 98, Martonchik 98 6 day repeat + over ocean AVn AVHRR,NOAA 81-90 Stowe 97 no land, a-priori - size+ abs- AVg AVHRR,GACP 84-00 Geogdzhyev 02 no land + cloud cont. Too TOMS 79-01 Torres 02 50km pixel, old rs ++ cloud cont Ton TOMS 79-01 Torres (priv.comm.) 50km pixel - surface refl POL POLDER 97,03 Deuzé 99, Deuzé 01 small size ov land + at high ele

3

Note, that ground-based monitoring is particular useful, if ground sites are part of a network with similar instrumentation and coordinated sampling. Examples are AERONET http://aeronet.gsfc.nasa.gov, SKYNETT

http://atmos.cr.chiba-u.ac.jp/, GAW http://wdca.jrc.it/ for passive remote sensing by sun-photometry and EARLINET http://www.earlinet.org/ for active remote sensing by lidar.

AOD data reference Monthly multi-annual data from ground-based sun-photometry are selected as

reference for an assessment and (regional) ranking of the different satellite AOD maps. The sun-photometer monthly statistics is based primarily on 1996-2006 AERONET data (Holben et al, 1998) further enhanced by 2003-2004 data of SKYNET (Aoki et al, 2006) and GAW. Seasonal AOD maps from sun-photometry are displayed in Figure 2.

Figure 2. Seasonal sun-photometer AOD maps of AERONET (1996-2006), SKYNET (2003-2004) and GAW (2003-2004) ground-based monitoring networks. In combining these local data on a global grid each site was associated with a quality weight and a regional representation weight - causing different domain (-disc) sizes. Local values were enlarged for better viewing

4

To simplify the assessments, individual station statistics of sun-photometer data

were combined onto the common 1o x 1o lat/lon grid of the satellite remote sensing data-sets. This gridding procedure considered known differences in quality and regional representation of the local data (T.Eck, personal communication). With the overall goal to combine identified regional retrievals strengths for a superior satellite AOD composite, assessments need to be conducted on a regional basis.

Regional stratification It is expected that a data-set that combines all regionally best performing satellite

retrievals will be superior in coverage and accuracy over any individual satellite retrieval. For the necessary regional stratification six oceanic and six continental zonal bands were chosen. The zonal bands separate Arctic, northern mid-latitudes, dust-belt, biomass belt, southern oceans and Antarctica, as illustrated in Figure 3. Figure 3 also compares zonal annual averages of the remote sensing data-sets, already introduced in Figure 1 and Table 2, separately over oceanic (left comparisons) and continental regions (right comparisons).

Figure 3. Regional choices and a comparison of mid-visible AOD of all remote sensing annual maps in Figure1. Averages of zonal bands over oceans are to the left and those over land are to the right. Regional averages follow the scale in the lower right and are only displayed, if satellite data had at least 25% or gridded sun-photometer data (AERONET) had at least 2.5% coverage.

In Figure 3 regional AOD averages are only displayed, if spatial coverage in that region exceeded 25% for satellite data or 2.5% for sun-photometer (1x1 gridded) data. In light of the different spatial sub-samples (see Figure 1) for satellite data over land and at high latitudes and certainly for sun-photometer statistics due to their local nature, these comparisons are more general in nature. More meaningful assessments of satellite AOD

5

data (to the sun-photometry reference) requires that all satellite data are sub-sampled in any region only at grid locations, where sun-photometer data exist.

Regional comparisons

Comparisons of matching data are presented in Figure 4 for regions, where non-zero (gridded 1o x 1o lat/lon) sun-photometer data (AERONET) cover at least 5% of that region. This limits satellite AOD data-set assessments to low- and mid-latitudes over land and only to northern hemispheric low- and mid-latitudes over oceans.

Figure 4. regional comparison of subsets of mid-visible AOD remote sensing data of Figure 1. All data-sets were sub-sampled only at grid-points with non-zero sun-photometer data and are only displayed, if the regional coverage of sun-photometer non-zero gridded data exceeded 5% in that region. Averages of zonal bands over oceans are to the left and those over land to the right.

This more visible evaluation indicates that in reference to ground based monitoring by sun-photometry almost all satellite retrievals overestimate AOD over oceans. Tendencies over land are more diverse and also a function of the dominant aerosol type (e.g. industrial pollution, desert dust or biomass burning). In comparison to sun-photometry, there is no single best retrieval for all regions. For the determination of a region’s superior satellite AOD retrieval a more objective rank scoring scheme was

6

applied to the satellite-retrieval vs. sun-photometry data pairs. The overall idea is to rank satellite data-sets as they compare locally to AERONET sun-photometer data and then use the results to create a merged data-set from satellite observations.

Scoring concept

In order to score the performance of global monthly data-sets with respect to a quality data reference, aspects of bias, spatial and seasonal variability should be included, while it is also desirable to summarize the overall performance by a single score. Thus a total score ST is defined (ST = sign(EB)* [1-|EB|]* [1-EV]* [1-ES]) with ‘0’ for poor and ‘1’ for perfect. This total score ST combines error-scores from a bias (EB), spatial variability (EV) and seasonality (ES). Moreover, the sign of ST reflects the bias with respect to the reference data. Error-scores (EB, EV and ES) are defined for a range from 0 for ‘perfect’ and 1 for ‘poor’. All error-scores are based on ranks (not values) to minimize the impact of data outliers. For the bias score EB involving N data-pairs from a data-set D and a reference data-set R, all 2*N elements are placed in one single array C and ranked by value (rank=1 for the lowest values rank=2*N for the largest value). Then the ranks of C associated with D and R are summed and compared. If the two sums of rank-values for D (DSUM) and R (RSUM) differ, then a bias and also its sign are identified by EB.

EB = w * [(DSUM-RSUM) / (DSUM+RSUM)], with w = [RANGE D+RANGE R]/[MEAN D+MEAN R]

The weight w is applied to avoid an overemphasis of this error in the overall score, in case all individual values are close to their average. The same weight is also applied for the (regional) variability score EV and the (temporal) seasonality score ES.

EV = w * [1-RC]/2 ES = w * [1-RC]/2, with w = [RANGE D+RANGE R]/[MEAN D+MEAN R]

Both variability scores are based on the (Spearman) rank correlation coefficient RC, which ranges from 1 for ‘correlated’ to -1 for ‘anti-correlated’. (Note that a lack of correlation [RC =0] still leads to a positive score for EV and ES).

Scores of any data-set D with respect to reference data-set R are first determined on a monthly and regional basis – considering regional variability and bias. Then these regional scores are combined with the seasonal variability score (using the monthly median data-pairs for D and R) into annual regional scores. Finally, these regional annual scores are combined according to their surface area fraction into annual global scores.

Global scores

Annual global scores (ST) for remote sensing AOD data of Figure 1 are presented in Table 3 based on regional and monthly sub-scores to AERONET sun-photometer statistics. Table 3 also provides contributing sub-scores for seasonality, bias and regional variability. In addition, also scores for AERONET sky-radiance data, a subset of sun-photometry is presented. Note, however, that due to different regional samples and coverage, the scores below do not represent a uniform test of satellite retrieval accuracy, which is beyond the scope or the current paper, but is subject of continuing work.

sky vs sun data

7

The AERONET sky-photometer (radiance) data are a subset of the (AERONET) sun-photometer (direct attenuation) data reference. Thus, better scores compared to satellite remote sensing are expected. The overall ‘sky’ score of -0.85 indicates through its negative sign that sky-data are biased low compared to sun-data and through the absolute value (0.85) that the match of monthly statistics between sky- and sun-data is not perfect (1.0). Sub-scores reveal that these deductions are mainly due to differences in spatial variability (0.94), which is a sampling problem, and due to the bias (-0.94).

Table 3. Global annual scores of remote sensing AOD data-sets versus the AERONET sun-photometer data reference (the higher the score to absolute one the better).

Rank label Overall score S T

season 1-E S

bias 1-E B

variability 1-E V

data-set

1 sky -.85 .96 -.94 .94 AERONET-sky 2 A,n .62 .90 .91 .76 AVHRR-NOAA *3 Mc5 .59 .94 .82 .78 MODIS coll. 5 4 Pol .55 .80 .89 .77 POLDER 5 A,g .51 .92 .74 .75 AVHRR-GACP* 6 Mc4 .51 .92 .73 .76 MODIS coll. 4 7 Mis .50 .87 .74 .77 MISR 8 T,n -.46 .80 -.93 .61 TOMS, new 9 T,o .38 .85 .66 .67 TOMS, old

note: AVHRR scores are based on ocean data only. AVHRR-NOAA data have a slight advantage as they were calibrated against the AERONET reference at selected sites. MISR data (with a narrow swath) andTOMS data (with a 50km*50km footprint) have fewer overall samples. And MISR and TOMS scores also include more difficult retrievals over bright desert surfaces, which are avoided in MODIS retrievals. Scores limited to either ocean or land regions are given in Tables 4 and 5. Table 4. Ocean annual scores for remote sensing AOD data vs the AERONET sun data reference

rank label Overall score S T

season 1-E S

bias 1-E B

Variability1-E V

data-set

1 sky -.81 .96 -.91 .93 AERONET-sky 2 A,n .62 .90 .91 .76 AVHRR-NOAA *3 Mc5 .58 .96 .76 .80 MODIS coll. 5 4 Mc4 .54 .94 .73 .78 MODIS coll. 4 5 Pol .55 .80 .89 .77 POLDER 6 A,g .51 .92 .74 .75 AVHRR-GACP* 7 Mis .49 .90 .70 .78 MISR 8 T,n .46 .77 .97 .62 TOMS, new 9 T,o .38 .86 .63 .69 TOMS, old

8

Table 5. Land annual scores for remote sensing AOD data vs the AERONET sun data reference

Rank Label Overall score S T

Season 1-E S

bias 1-E B

Variability1-E V

data-set

1 sky .91 .97 .98 .95 AERONET-sky 2 Mc5 .61 .88 .95 .73 MODIS coll. 5 3 Pol .56 .76 .99 .75 POLDER 4 Mis .54 .86 .84 .75 MISR 5 Mc4 .45 .88 .73 .70 MODIS coll. 4 6 T,o .37 .83 .73 .62 TOMS, old 7 T,n -.36 .88 -.70 .58 TOMS, new

Interestingly, separate scores for oceanic and continental regions in Table 4 and

Table 5 indicate that the negative AOD bias of the sky-data occurs in oceanic regions, while the AOD bias of sky-data is positive over land. This indicates, that aside from the more conservative cloud-screening by sky data (which would result in lower sky-AOD values) other factors such as sampling may have impacts on monthly averages as well. satellite vs sun-data

The overall scores for the satellite AOD retrievals range from 0.62 (better) to 0.38 (poorer). Almost all satellite remote sensing data display positive overall scores. Thus, if AERONET sun-photometer AOD data are trusted, this indicates that (on average) almost all satellite retrievals overestimate AOD. Sub-scores indicate that the largest deductions are usually associated with spatial variability and also with the bias in some data-sets. Scores for regional and temporal variability are poorer over land than over oceans. Updated retrievals of the same sensor data (e.g. MODIS, TOMS) show improved overall scores. This indicates the untouched potential of sensor data und the need for continued retrieval improvements.

In order to create a satellite composite the strongest performing retrievals on a regional basis must be identified. Scores for all (except polar) regions are listed for the continental zonal latitude bands in Table 6 and for the oceanic zonal latitude bands in Table 7. As a reference, also the scores of the model median AOD fields of global model simulations by AeroCom (Kinne et al., 2006) are provided. When comparing scores, however, it should be kept in mind that the sample volume differs among data-sets. It is possible that poorer sampling for MISR and TOMS contributed to their lower scores. Thus, unless the sampling impact is better understood, scores should not be interpreted as retrieval error. Also the scoring is only based on AERONET locations, where satellite data are available (e.g. MODIS does not provide data over bright land surfaces), which also explains differences for the reference median (RMEDIAN), especially over land.

Table 6. Oceanic regional annual scoring and sub-scoring of remote sensing AOD data-sets with respect to AERONET sun-photometer data. Scores of median fields by global modeling are added.

9

rank score(ST) 1-ES 1-EB 1-EV DMEDIAN RMEDIAN 58N-30N ocean 1 -0.9166 0.990 -.963 0.961 0.122 0.126 AERONET-sky 2 0.6610 0.928 0.883 0.807 0.148 0.120 POLDER 3 0.6209 0.870 0.818 0.873 0.179 0.120 MODIS col.5 4 0.5952 0.895 0.839 0.793 0.167 0.120 AVHRR-GACP 5 0.5948 0.862 0.852 0.810 0.149 0.109 AVHRR-NOAA 6 0.5911 0.909 0.753 0.864 0.209 0.120 MODIS col.4 7 0.5378 0.874 0.754 0.816 0.205 0.120 MISR 8 0.3860 0.874 0.669 0.661 0.271 0.129 TOMS old

9 -0.3770 0.957 -.662 0.595 0.090 0.120 TOMS new M 0.6713 0.950 0.845 0.836 0.173 0.120 model median

rank score(ST) 1-ES 1-EB 1-EV DMEDIAN RMEDIAN 30N-10N ocean 1 -0.8829 0.974 -.931 0.974 0.099 0.103 AERONET-sky 2 0.7961 0.941 0.973 0.870 0.110 0.103 AVHRR-NOAA 3 0.7450 0.991 0.830 0.906 0.154 0.103 MODIS col.5 4 0.7335 0.995 0.817 0.903 0.162 0.103 MODIS col.4 5 0.7180 0.939 0.858 0.892 0.149 0.103 POLDER 6 0.7097 0.967 0.807 0.910 0.163 0.103 MISR 7 0.6344 0.843 0.874 0.862 0.137 0.103 AVHRR-GACP 8 -0.6006 0.940 -.945 0.676 0.105 0.103 TOMS new

9 0.5525 0.874 0.747 0.847 0.208 0.103 TOMS old M 0.7931 0.901 0.980 0.898 0.112 0.103 model median

rank score(ST) 1-ES 1-EB 1-EV DMEDIAN RMEDIAN 10N-22S ocean 1 -0.8399 0.969 -.988 0.878 0.083 0.071 AVHRR-NOAA 2 -0.7730 0.944 -.879 0.932 0.061 0.071 AERONET-sky 3 0.7136 0.978 0.816 0.894 0.112 0.071 MODIS col.5 4 0.6953 0.927 0.846 0.886 0.104 0.071 MODIS col.4 5 0.6715 0.971 0.824 0.839 0.124 0.071 AVHRR-GACP 6 0.6099 0.925 0.743 0.888 0.138 0.071 MISR 7 0.5494 0.893 0.768 0.800 0.158 0.071 POLDER 8 0.4598 0.720 0.911 0.700 0.112 0.071 TOMS new

9 0.4387 0.856 0.661 0.775 0.199 0.071 TOMS old M -0.5602 0.808 -.794 0.873 0.056 0.079 model median

rank score(ST) 1-ES 1-EB 1-EV DMEDIAN RMEDIAN 22S-58S ocean 1 -0.7517 0.942 -.888 0.899 0.046 0.057 AERONET-sky 2 0.4238 0.842 0.831 0.606 0.084 0.073 AVHRR-NOAA 3 0.3737 0.959 0.643 0.605 0.120 0.073 MODIS col.5 4 0.3545 0.587 0.885 0.683 0.076 0.073 POLDER 5 0.3433 0.929 0.586 0.631 0.130 0.073 AVHRR-GACP 6 0.3103 0.933 0.568 0.586 0.126 0.073 MODIS col.4 7 0.2810 0.653 0.828 0.520 0.085 0.073 TOMS new 8 0.2770 0.839 0.571 0.579 0.136 0.073 MISR

9 0.2416 0.850 0.533 0.533 0.204 0.068 TOMS old M 0.3511 0.872 0.808 0.498 0.088 0.073 model median

10

Table 7. land regional annual scoring and sub-scoring of remote sensing AOD data-sets with respect to AERONET sun-photometer data. Scores of median fields by global modeling are added.

rank score(ST) 1-ES 1-EB 1-EV DMEDIAN RMEDIAN 58N-30N land 1 0.9214 0.975 0.983 0.961 0.127 0.120 AERONET-sky 2 0.5612 0.847 0.823 0.805 0.179 0.121 MISR 3 0.5337 0.749 0.969 0.736 0.141 0.136 POLDER 4 0.5077 0.848 0.858 0.699 0.171 0.121 MODIS col.5 5 0.4319 0.953 0.662 0.685 0.268 0.120 TOMS old 6 0.4048 0.886 0.660 0.692 0.255 0.122 MODIS col.4

7 -0.3109 0.972 -.623 0.514 0.046 0.121 TOMS new M 0.7158 0.974 0.888 0.828 0.156 0.121 model median

rank score(ST) 1-ES 1-EB 1-EV DMEDIAN RMEDIAN 30N-10N land 1 0.9291 0.988 0.975 0.965 0.371 0.359 AERONET-sky 2 -0.6697 0.925 -.975 0.743 0.407 0.361 MISR 3 -0.6377 0.913 -.882 0.792 0.321 0.367 MODIS col.5 4 0.6216 0.939 0.902 0.734 0.458 0.383 MODIS col.4 5 -0.5398 0.732 -.900 0.819 0.269 0.301 POLDER 6 0.5183 0.836 0.996 0.622 0.376 0.361 TOMS old

7 -0.4520 0.893 -.724 0.700 0.208 0.366 TOMS new M -0.5844 0.869 -.864 0.778 0.309 0.361 model median

rank score(ST) 1-ES 1-EB 1-EV DMEDIAN RMEDIAN 10N-22S land 1 -0.8568 0.986 -.959 0.907 0.119 0.134 AERONET-sky 2 -0.7260 0.958 -.990 0.765 0.156 0.140 MODIS col.5 3 0.6937 0.955 0.963 0.754 0.115 0.110 POLDER 4 0.5057 0.850 0.825 0.722 0.269 0.138 MISR 5 0.4829 0.885 0.774 0.705 0.254 0.138 MODIS col.4 6 0.3599 0.713 0.760 0.664 0.301 0.138 TOMS old

7 -0.3485 0.725 -.737 0.653 0.080 0.139 TOMS new M -0.5338 0.905 -.796 0.742 0.098 0.138 model median

rank score(ST) 1-ES 1-EB 1-EV DMEDIAN RMEDIAN 22S-58S Land 1 -0.9556 0.989 -.996 0.969 0.076 0.080 AERONET-sky 2 0.6670 0.913 0.994 0.736 0.073 0.068 MODIS col.5 3 0.6261 0.974 0.776 0.828 0.116 0.068 MISR 4 0.4446 0.951 0.653 0.716 0.186 0.068 MODIS col.4 5 0.4336 0.820 0.921 0.575 0.079 0.068 TOMS new 6 0.3203 0.795 0.623 0.647 0.181 0.068 TOMS old

7 0.1127 0.188 0.864 0.694 0.102 0.075 POLDER M 0.5249 0.787 0.906 0.736 0.078 0.068 model median

No individual satellite retrieval displays the highest score. Thus, a satellite

composite was created by combining the regionally highest scoring retrievals.

11

Satellite composite

Over continents the highest scores are achieved by MISR for northern

hemispheric urban pollution and for dust and by MODIS (collection 5) for southern hemispheric biomass burning. Over oceans the highest scores are by POLDER for the higher latitudes of the northern hemisphere and AVHRR-NOAA over all other oceans (In fairness to more advanced recent retrievals over oceans it should be admitted that the AVHRR-NOAA retrieval used sun-photometry for sensor calibration). To reduce the potential for abrupt changes when switching between different retrievals at zonal band boundaries and land/ocean transitions, a 6 degree latitudinal transition zone and smoothing in coastal regions was permitted. The annual global AOD maps of the composite and its major contributing retrievals are presented in Figure 5.

Figure 5. Comparison of the annual mid-visible AOD maps of the satellite composite (co) to contributing multi-annual data of MISR (MIS), MODIS coll. 5 (Mc5) and AVHRR-NOAA (AVn).

12

Seasonal AOD maps of the composite are presented in Figure 6. They illustrate that AOD maxima due to biomass burning and dust have a strong seasonal character, which gets lost in the presentation of annual global maps.

Figure 6. Seasonal (mid-visible) AOD maps for the satellite composite Although retrievals with the regional highest scores with respect to sun-photometer statistics are applied none of these retrievals scores close to perfect. Thus, differences to the reference data can be expected and are illustrated on a seasonal basis in Figure 7.

13

Figure 7. Seasonal AOD differences between the satellite composite and (sparse and visually enlarged) AERONET site statistics. Positive differences indicate overestimates and negative values indicate underestimates. Only AOD differences exceeding +/-0.05 are displayed. The AOD satellite composite still displays strong differences to the reference from sun-photometry. Higher AOD values in northern mid-latitudes during winter and spring suggest snow-contamination. Lower AOD values near fast growing urban pollution areas and dust outflow regions seem to suggest that some of the larger AOD events are missed due to event removals in cloud-screening algorithms. The same reason may also be responsible for the most severe deviations as the timing of the tropical biomass maxima occurs too early (in late summer) compared to the sun-photometer reference (early autumn). To reduce these deviations in a measurement based AOD composite, a modified version of the satellite composite has been developed, where the monthly AOD fields of the composite are drawn to the AERONET data.

14

Enhanced composite

Information of sun-photometer monthly statistics was added to the AOD satellite composite. This merging of data gave priority to sun-photometer statistics when locally available. First background ratio field are defined by globally spreading available sun-photometer to satellite composite ratios with distance decaying weights. Then these ratios are applied in site surrounding domains. Global annual (mid-visible) AOD maps of this new enhanced satellite composite in comparison to contributing fields of the original satellite composite and of sun-photometry (AERONET) samples are presented in Figure 8. For comparison the annual AOD map from global modeling median is given as well.

Figure 8. Annual mid-visible AOD maps of the sun-photometry enhanced satellite composite (cR) and its contributing maps of the satellite composite (co) and AERONET (aer). For a comparison the annual global AOD map based on local monthly median values from global modeling is displayed (med) as well. The new enhanced AOD composite had modified the initial satellite composite mainly over the western part of North America and over the tropical biomass regions of

15

South America. Seasonal maps of the (new) enhanced AOD composite are presented in Figure 9.

Figure 9. Seasonal (mid-visible) AOD maps of the AERONET enhanced satellite composite The comparison of seasonal AOD maps between the new (Figure 9) and the initial (Figure 6) satellite composites in Figure 6, demonstrates for the enhanced composite reduced AOD values during continental winters of the northern mid-latitudes and increased AOD values near tropical biomass burning regions and urban pollution in Asia during fall. Remaining differences of the enhanced composite to the sun-photometer reference are illustrated on a seasonal basis in Figure 10.

16

Figure 10. Seasonal AOD differences between the AERONET enhanced satellite composite and (sparse and visually enlarged) AERONET site statistics. Positive deviations show overestimates and negative values show underestimates. Only AOD differences exceeding +/-0.05 are shown. Annual global averages for deviations are listed below the labels. The differences of the (new) enhanced composite to the AERONET reference (Figure 10) have been reduced compared to those of the satellite composite (Figure 7). Some larger deviations near urban centers (e.g. Mexico City, East Asia) or biomass sources (e.g. central South America) remain, as the local influence on surrounding regions was apparently overextended in the applied data-merging. With improved merging procedures a closer fit to AERONET (and smaller deviations) can be expected. As the (new) enhanced composite displays overall smaller deviations to the reference, quantitative scores should have significantly improved. And they do, as illustrated in Tables 8, 9 and 10 for global, oceanic and continental scores. The sub-scores demonstrate that the biggest improvements are to the temporal and the spatial variability.

17

Table 8. Global annual scoring of AOD data composites and median fields of global modeling

Rank label overall score S T

season 1-E S

bias 1-E B

variability 1-E V

data-set

1 cR .73 .96 .92 .82 enhanced comp 2 co .65 .92 .91 .78 sat composite 3 med .63 .85 .99 .75 global modeling

Table 9. Ocean annual scoring of AOD data composites and median fields of global modeling

Rank label overall score S T

season 1-E S

bias 1-E B

variability 1-E V

data-set

1 cR .72 .97 .92 .82 enhanced comp 2 co .66 .93 .91 .78 sat composite 3 med .62 .85 .98 .74 global modeling

Table 10. Land annual scoring of AOD data composites and median fields of global modeling

rank label overall score S T

season 1-E S

bias 1-E B

variability 1-E V

data-set

1 cR .74 .94 .93 .85 enhanced comp 2 med -.64 .84 -.98 .77 global modeling 3 co .61 .89 .90 .76 sat composite

global modeling

An alternative source for global AOD maps is the use of simulated distributions from global modeling. For characteristic AOD maps from global modeling, a composite of monthly local median values of simulations with twenty different models of AeroCom exercises was chosen (Kinne et al. 2006). The scores of the model median data-set rank with the best data-sets from remote sensing and are comparable to those of the composite.

Median data of an ensemble reduce the impact of outliers by individual models. Thus, relatively good scores can be expected which are at least superior to the ensemble average. Nonetheless, there are apparent biases of this model median data-set compared to regional and seasonal distribution suggested by remote sensing. Based on the regional (bias) scores in Tables 6 and 7, global modeling tends to overestimate AOD over both land and ocean at northern mid-latitudes (outdated emission inventories?) and tends to underestimate AOD over land regions with major dust and biomass burning sources and over remote oceanic regions of the Southern hemisphere. These general modeling biases are illustrated in differences to the (data-based) enhanced composite on a seasonal basis in Figure 11.

18

Figure 11. Seasonal AOD differences between the sun-photometry enhanced (satellite) composite and median fields of global modeling. Positive differences suggest model underestimates, while negative deviations indicate model overestimates. Only AOD differences exceeding +/-0.05 in absolute value are displayed. Annual global averages for deviations are listed below the labels.

The deviations are significant, as these are not individual events but seasonal

averages. The (relative) low bias in modeling over remote oceans does not show, as Figure 11 addresses absolute errors, while (mid-visible) AOD values in that remote oceanic regions based on sun-photometry are on the order of 0.06.

Conclusion

When in need for climatological data on aerosol, there is a temptation to adopt data from global modeling. The advantage of global modeling is that the data are complete (e.g. no data gaps) and consistent (e.g. all aerosol properties). However, there should be awareness that data, and in particular aerosol data, from global modeling are based on often poorly constrained interactions and input data. More specifically, despite the ability to distinguish between different aerosol types in advanced aerosol modules,

19

the required aerosol input data (e.g. emission source strength and location) and the aerosol process parameterizations (e.g. transport and removal) are highly uncertain. In fact, the resulting AOD maps in the end are usually tuned towards available AOD data from observations. Thus, to take advantage of the extra detail by global modeling (e.g. information on properties that cannot be measured) there is a demand for quality constrains by measurement based data. In the case of aerosol, reliable global monthly maps for (mid-visible) AOD (and AOD solar spectral dependence) would be extremely useful. Unfortunately, individual data sources usually lack either spatial coverage (e.g. sun-photometry) or accuracy (e.g. satellite remote sensing). Thus, methods need to be explored that combine the strength of individual data sources in order to create a superior data product that can really help modeling. The enhanced AOD composite developed in this contribution is far from perfect. But it demonstrates that there are ways to more useful data-products. Thus, as an incentive to continue on that path, Figure 12 displays monthly global maps of the enhanced AOD composite from remote sensing.

Figure 12. Monthly global maps for the mid-visible aerosol optical depth based on an AERONET enhanced satellite retrievals composite.

20

References Aoki, K., Aerosol optical characteristics in Asia from measurements of SKYNET sky radiometers. in IRS

2004: Current Problems in Atmospheric Radiation, H. Fischer and B-J. Sohn, Eds., A. Deepak Publishing, 311-313 (2006).

Deuzé,J.L, F.M. Breon, C. Devaux, P. Goloub, M. Herman, B. Lafrance, F. Maignan, A. Marchand, F.

Nadal, G. Perry and D.Tanré: Remote sensing of aerosol over land surfaces from POLDER/ADEOS-1 polarized measurements, J.Geophys.Res., 106, 4912-4926, 2001.

Geogdzhyev, I., M.Mishchenko, W.Rossow, B.Cairns and A.Lacis: Global 2-channel AVHRR retrieval of

aerosol properties over the ocean for the period of NOAA-9 observations and preliminary retrievals using NOAA-7 and NOAA-11 data, J.Atmos.Sci., 59, 262-278, 2002.

Holben, B., T.Eck, I.Slutsker, D.Tanre, J.Buis, E.Vermote, J.Reagan, Y.Kaufman, T.Nakajima, F.Lavenau,

I.Jankowiak and A.Smirnov: AERONET, a federated instrument network and data-archive for aerosol characterization, Rem.Sens.Environ., 66, 1-66, 1998.

Kahn, R., P.Banerjee, D.McDonald and D.J.Diner: Sensitivity of multi-angle imaging to aerosol optical

depth and to pure particle size distribution and composition over ocean, J.Geophys.Res., 103, 32195-32213, 1998.

Kaufman, Y., D.Tanré, L.Remer, E.Vermote, D.Chu and B.Holben: Operational remote sensing of

tropospheric aerosol over the land from EOS-MODIS, J.Geophys.Res., 102, 17,051-17,061, 1997. Kinne, S., M.Schulz, C.Textor, S.Guibert, Y.Balkanski, S.Bauer, T.Berntsen, T.Berglen, O.Boucher, M.

Chin, F.Dentener, T.Diehl, H.Feichter, D.Fillmore, S.Ghan, P.Ginoux, S.Gong, A.Grini, J.Hendricks, L.Horowitz, I.Isaksen, T.Iversen, A.Kirkevåg, S.Kloster, D.Koch, J.E.Kristjansson, M.Krol, A.Lauer, J.F.Lamarque, X.Liu, V.Montanaro, G.Myhre, J.Penner, G.Pitari, S.Reddy, Ø. Seland, P.Stier, T.Takemura and X.Tie : An AeroCom initial assessment – optical properties in aerosol component modules of global models. Atmos.Chem.Phys., 6, 1815-1834, 2006.

Liu, L.and Mishchenko, M.I., 2008: Toward unified satellite climatology of aerosol properties: direct

comparisons of advanced level 2 aerosol products, J. Quant. Spectr. Rad. Transfer, in press. Martonchik J.V., D.J.Diner, R.A.Kahn, T.P.Ackerman, M.E.Verstraete, B.Pinty and H.R.Gordon:

Techniques for the retrieval of aerosol properties over land and ocean using multi-angle imaging, IEEE Trans. Geosci. Remote Sensing 36, 1212-1227, 1998.

Schulz, M., C.Textor, S.Kinne, S.Guibert, Y.Balkanski, S.Bauer, T.Berntsen, T.Berglen, O.Boucher, M.

Chin, F.Dentener, T.Diehl, H.Feichter, D.Fillmore, S.Ghan, P.Ginoux, S.Gong, A.Grini, J.Hendricks, L.Horowitz, I.Isaksen, T.Iversen, A.Kirkevåg, S.Kloster, D.Koch, J.E.Kristjansson, M.Krol, A.Lauer, J.F.Lamarque, X.Liu, V.Montanaro, G.Myhre, J.Penner, G.Pitari, S.Reddy, Ø. Seland, P.Stier, T.Takemura and X.Tie : Radiative forcing by aerosols as derived from the AeroCom present-day and pre-industrial simulations. Atmos.Chem.Phys.Disc., 6, 5095-5136, 6 2006.

Stowe L., A.Ignatov and R.Singh: Development, validation and potential enhancements to the second

generation operational aerosol product at NOAA/NESDIS, J.Geophys.Res. 102, No. D14, pp. 16923-16934, 1997.

Tanré, D., Y.J.Kaufman, M.Herman and S.Mattoo, Remote Sensing of aerosol properties over ocean using

the MODIS/EOS spectral radiances, J. Geophys. Res., 102, 16,971-16,988, 1997.

21

Torres, O., P.K.Barthia, J.R.Herman, A.Sinyuk, P.Ginoux and B.Holben: A long-term record of aerosol optical depth from TOMS observationsand comparisons to AERONET measurements. J.Atmos.Sci. 59, 398-413, 2002.

22