QA filtering of individual pixels to enable a more accurate validation of aerosol products Maksym Petrenko Presented at MODIS Collection 7 and beyond Retreat

MAPSS supports Level 2 aerosol data from different sensors – MODIS – MISR – OMI – POLDER – CALIOP – AERONET MAPSS uniformly samples Level-2 aerosol products from multiple sensors over uniform areas of 55km centered around AERONET sun photometer ground stations Stores resulting statistics in simple CSV files MAPSS: Multi-sensor Aerosol Products Sampling System

Sensor ground footprint (at nadir) 50 km 11/16/113

Subset statistics General – Number of pixels in sample space (Ndat) – Valid pixel count (Nval) – Closest pixel value (Cval) – Mean, Median, Mode – Standard Deviation Spatio-temporal variability – Slope of fitted Plane or Line – Azimuth (direction) of slope – Multiple (or Linear) Correlation Coefficient 11/16/114 Ndat=9 Nval =6

Additional subset data Geometry – Solar zenith angle – Sensor zenith angle – Scattering angle – … and so on Data provenance – Name of source data file – Index of the closest pixel in the data file Quality control / Quality assurance (QA) – Mode or mean of QA flags in the subset area – Bit mask flags are decoded and reported as plain numbers 11/16/115

Accounting for data quality OMI data with all QAOMI data with best QA MODIS data with all QAMODIS data with best QA 11/16/116

QA experiment Compare means of satellite-derived AOD over 55km areas to means of AERONET AOD (interpolated) within ±30 minutes of satellite overpasses Filter satellite data based on QA: – No filtering (left column) – Compute mean AOD based only on pixels with Best QA (central column) – Compute mean AOD based on all pixels in the sample, but only if mode of QA over the whole sample is Best (right column)

MODIS AOD Land - Corrected No QA filtering Pixel QA=3 Mode QA=3

MODIS AOD Ocean Average No QA filtering Pixel QA=3 Mode QA=3

MODIS AOD Deep Blue No QA filtering Pixel QA=3 Mode QA=3

OMI No QA filtering Pixel QA=0 Mode QA=0

Observations For MODIS products, considering the overall data quality in the observation region produces a more accurate (but smaller) subset of the data Applying QA filtering to individual pixels in MODIS Land or Ocean products discards certain data points, but does not change the overall character of the data For OMI, pixel-based filtering produces somewhat better results, perhaps because of higher variability of measurement error in sampling areas (e.g., subpixel cloud contamination)

Dust storm at Djougou on February 26, 2005 Source: MODIS Terra Rapid Response, AERONET Data Synergy Tool 14

Daily AERONET AOD data during FEB 2005 at Djougou 15 Source: AERONET site

Best-QA mean AOD from multiple sensors at Djougou The event was observed by multiple sensors, but most best-QA measurements underestimated the actual AOD value Interestingly, there were no best-QA Deep Blue AOD retrievals from MODIS, but there were some from SeaWiFS 16

Dark Target and Deep Blue AOD from MODIS at Djougou AOD at Djougou is retrieved by MODIS using both Dark Target and Deep Blue algorithms However, only Dark Target retrievals have the best quality 17

All-QA AOD from multiple sensors at Djougou 18 A significantly larger number of lower-QA observations indicates the conditions (i.e., bright surface, high aerosol loading) were hard for most sensors Deep Blue AOD from MODIS and SeaWiFS seem to be the most accurate in these conditions, however they should be used with caution because of the low QA AOD retrieved by OMI on Feb.26 was the closest to the peak AERONET AOD, but was also indicated as having a low QA

Mean Angstrom Exponent at Djougou 19

Mean Single Scattering Albedo at Djougou 20

Potential of QA misrepresentation Under extreme or uncertain conditions, retrieval algorithms might have difficulty assigning the correct QA flags, and even valid data might be marked as “Bad” Analysis of long-term data can help to reveal areas with systematic uncertainties in QA values, e.g. coastal areas For example at COVE region (coast line of Virginia), most of the sensors retrieve AOD in a close agreement to AERONET (left plot). However, only a small portion of these data has Best QA (right plot, please note a changed scale)

Summary During validation studies, filtering individual pixels by QA and then computing mean, rather than filtering the mean of the sample by the QA mode, might produce more accurate (albeit somewhat worse) results Bad-QA data with good correlation to AERONET should be investigated for a possibility of raising its score