440 likes | 465 Views
Explore implementation strategies and inversion methods for satellite remote sensing of Inherent Optical Properties (IOPs) at Ocean Optics XIX workshop in Oct. 2008 moderated by Bryan Franz from NASA/OBPG.
E N D
Topic 3 Operational Implementation Strategies Moderator: Bryan Franz IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Goals • Identification of issues unique to satellite retrieval of IOPs • Understanding of satellite Rrs generation • Agreement on common IOP model inputs (aw & bbw) • Agreement on algorithm failure conditions & masking • Understanding impact of IOP inversion at L2 versus L3 IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Satellite Focus • Multiple sensors - varying wavelength sets • SeaWIFS, MODIS, MERIS --> OCM-2, VIIRS, OLCI • Multiple data processing systems (NASA, ESA, ISRO) • Global application • wide range of water classes, distribution dominated by low-Ca water • large data volumes, want best IOP algorithm that is “practical” • Imperfect Lw retrieval • satellite sensor calibration & noise • atmospheric correction error • Rrs normalization • wide range of viewing geometry (0 < v < 60) • transition through interface IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Multi-Sensor Processing Framework atmospheric correction Lw normalization derived products flags (failure & quality) Level-1 to Level-2 (common algorithms) observed Lt() , 0, SeaWiFS L1A - or - MODISA L1B MODIST L1B OCTS L1A MOS L1B OSMI L1A CZCS L1A MERIS L1B OCM L1B observed radiances AOPs Rrs() IOPs a(), bb() Level-2 File Level-2 File Level-2 File spatial averaging temporal averaging masking Level-2 to Level-3 Level-3 Global Product IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
full-band water-leaving radiance normalized to non-attenuating atmosphere with Sun overhead Sun fb nLw() = Lw() / td0() 0 f0 fb correct from full-band to nominal 10-nm center-band via Morel model nLw() = nLw() f correct for Fresnel reflection refraction and inhomogeneity of subsurface light field via LUT 0 (f/Q)0 ex nLw() = nLw() (f/Q) solar irradiance from Thuillier 2002 10-nm square-band-pass average ex ex Rrs() = nLw() / F0 () Rrs from Satellite Radiances TOA gas pol glint whitecap air aerosol td() Lw() = Lt() / tg() /fp() - TLg()- tLf() - Lr() - La() IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
IOP Model Implementation Issues • transition across air/sea interface • Lee et al. 2002 • pure sea-water values (aw & bbw) • aw: Pope & Fry, Kou et al. 1993, bbw: Smith & Baker 1981 • 10-nm square band-pass average (consistent with Rrs retrieval) • salinity & temperature sensitivity • significant impact on IOP retrieval when aw & bbw = f(T,S) • need to identify ancillary data sources IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Inversion Methods and Efficiency • sequential (class 1 & 3) • model-specific (wavelength-specific) • may be iterative • simultaneous (class 2) • Matrix inversion • Lower-Upper Decomposition (LUD) • Singular-Valued Decomposition (SVD) • Iterative cost-function minimization • Levenburg-Marquart (LM) • Downhill Simplex (Amoeba, AMB) Algorithm Time (secs) one SeaWIFS GAC orbit IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
IOP Models Implemented at NASA • GSM (Garver-Siegel-Maritorena) • a, aph, adg, bb, bbp, Ca • QAA (Quasi-Analytical Algorithm) • a, aph, adg, bb, bbp • LAS (Loisel and Stramski) • a, b, c, bb, bbp • PML (Plymouth Marine Labs) • a, aph, adg, bb, bbp • HAL (Hoge & Lyon, via GIOP) • a, aph, adg, bb, bbp • GIOP (Generalized IOP Model) • a, aph, adg, bb, bbp, Ca, flags, , S • TBD: • NIWA • Boss & Roesler IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
specify sensor wavelengths to fit e.g., 412,443,490,510,555 e.g., 412,490,555 select aph form and set params tabulated: , ap*() gaussian: , select adg form and set params exponential: , S select bbp form and set params power law: , power law: , via Hoge & Lyon power law: , via QAA select Rrs[0-] to bb/(a+bb) quadratic: g1, g2 f/Q: (tbd) specify inversion method Levenburg-Marquart Amoeba (downhill simplex) Lower-Upper Decomposition Singular-Value Decomposition specify output products a (), aph (), adg (), bb (), bbp () = any sensor wavelength(s) Ca (given ap* at ) (dynamic model params) internal flags Generalized IOP Model (GIOP) IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
specify sensor wavelengths to fit e.g., 412,443,490,510,555 e.g., 412,490,555 select aph form and set params tabulated: , ap*() gaussian: , select adg form and set params exponential: , S select bbp form and set params power law: , power law: , via Hoge & Lyon power law: , via QAA select Rrs[0-] to bb/(a+bb) quadtratic: g1, g2 f/Q: (tbd) specify inversion method Levenburg-Marquart Amoeba (downhill simplex) Lower-Upper Decomposition Singular-Value Decomposition specify output products a (), aph (), adg (), bb (), bbp () = any sensor wavelength(s) Ca (given ap* at ) (dynamic model params) internal flags Generalized IOP Model (GIOP) 5-Band GSM IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
specify sensor wavelengths to fit e.g., 412,443,490,510,555 e.g., 412,490,555 select aph form and set params tabulated: , ap*() gaussian: , select adg form and set params exponential: , S select bbp form and set params power law: , power law: , via Hoge & Lyon power law: , via QAA select Rrs[0-] to bb/(a+bb) quadratic: g1, g2 f/Q: (tbd) specify inversion method Levenburg-Marquart Amoeba (downhill simplex) Lower-Upper Decomposition Singular-Value Decomposition specify output products a (), aph (), adg (), bb (), bbp () = any sensor wavelength(s) Ca (given ap* at ) (dynamic model params) internal flags Generalized IOP Model (GIOP) Hoge & Lyon IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Flags & Masks IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Multi-Sensor Processing Framework atmospheric correction Lw normalization derived products flags (failure & quality) Level-1 to Level-2 (common algorithms) observed Lt() , 0, SeaWiFS L1A - or - MODISA L1B MODIST L1B OCTS L1A MOS L1B OSMI L1A CZCS L1A MERIS L1B OCM L1B observed radiances AOPs Rrs() IOPs a(), bb() Level-2 File Level-2 File Level-2 File spatial averaging temporal averaging masking Level-2 to Level-3 Level-3 Global Product IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Level-2 Flags & Level-3 Masking Level-2 flags used as masks in Level-3 processing IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Proposed Conditions for IOP Product Failure • Rrs < 0 in any required band? • not required for Rrs() minimization • required for matrix inversion (no positive roots in Gordon quad.) • required for band ratio component algorithms (e.g., QAA, HAL) • Failure within model computation • e.g., inputs out of range of LUTs, divide by zero errors • Tests on IOP retrievals (for 400 < < 600) or Rrs < -() initial proposal employed in some of our global analyses for this workshop IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
L2 vs L3 Inversion IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
IOP Inversion at Level-2 Standard NASA Approach atmospheric correction Lw normalization derived products flags (failure & quality) Level-1 to Level-2 (common algorithms) observed Lt() , 0, SeaWiFS L1A - or - MODISA L1B MODIST L1B OCTS L1A MOS L1B OSMI L1A CZCS L1A MERIS L1B OCM L1B observed radiances AOPs Rrs() IOPs a(), bb() Level-2 File Level-2 File Level-2 File Level-2 to Level-3 Level-3 Global Product IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
IOP Inversion at Level-3 Alternative Approach atmospheric correction Lw normalization derived products flags (failure & quality) Level-1 to Level-2 (common algorithms) observed Lt() , 0, SeaWiFS L1A - or - MODISA L1B MODIST L1B OCTS L1A MOS L1B OSMI L1A CZCS L1A MERIS L1B OCM L1B AOPs Rrs() IOPs a(), bb() averaged Rrs() , 0, Level-2 File Level-2 File Level-2 File Level-3 Global Product Level-2 to Level-3 Level-3 Global Product IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
L2 vs L3 Inversion: GSM Model IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
L2 vs L3 Inversion: GSM Model bb() < 0.015 mask GSM: Largest differences in eutrophic bb. IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
L2 vs L3 Inversion: QAA Model QAA: Largest differences in eutrophic a (band ratio algorithm, mean of ratio not same as ratio of means). IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
L2 vs L3 Inversion: PML Model Eutrophic Mesotrophic Oligotrophic a(443) bb(443) PML: differences everywhere (f/Q from mean geometry) IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Model-to-Model Differences Mesotrophic Oligotrophic Eutrophic a(443) bb(443) IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
My View • I like simultaneous solutions (class-1) • take advantage of full spectral suite, readily adapted to multiple sensors, easy to incorporate new ideas or alternative basis functions. • I prefer Rrs minimization to matrix inversion • can handle negative Rrs (small Rrs +/- noise) • seems less sensitive to noise (perhaps a weighting issue) • Efficiency in algorithm/inversion selection is not a primary concern • satellite data processing is i/o intensive (exception Boss & Roesler) • Inversion at Level-3 vs Level-2 is not a primary concern • differences between popular models are much greater • Mask all IOP products at Level-3 if: • any one product exceeds valid (TBD) range IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Discuss ... IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Goals • Identification of issues unique to satellite retrieval of IOPs • Understanding of satellite Rrs generation • Agreement on common IOP model inputs (aw & bbw) • Agreement on algorithm failure conditions & masking • Understanding impact of IOP inversion at L2 versus L3 IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Inversion Method IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Inversion Methods • sequential • model-specific • may be iterative • simultaneous • Iterative cost-function minimization • Levenburg-Marquart (LM) • Downhill Simplex (Amoeba, AMB) • Matrix inversion • Lower-Upper Decomposition (LUD) • Singular-Valued Decomposition (SVD) A x = b IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Rrs Minimization vs Matrix Inversion a(443), 6-Band GSM Model, SVD Fit a(443), 6-Band GSM Model, LM Fit a(443), 5-Band GSM Model, SVD Fit a(443), 5-Band GSM Model, LM Fit 5-Band = 412,443,490,510,555 a(443) 0.01 1.0 IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG 6-Band = 5-Band + 670
Matrix Inversion: Linearization Issue Rrs[0-] = g1 u + g2 u2 where u bb/(a+bb) - or - Rrs[0-] = f/Q u 1) Traditional Approach: System of Equations Proportional to 1/Rrs where v = 1 - 1/u a = -v bb aph() + adg ( + v bbp() = -[aw() + v bbw()] 2) Alternate Approach: System of Equations Proportional to Rrs u a = (1 - u) bb u aph() + u adg ( + (u-1) bbp() = -[u aw() + (u-1) bbw()] IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Alternate Linearization Improves Inversion Consistency Linearization Method 1 a(443) GSM 6-Band a(443) Global bb(443) Global LM LM SVD SVD Linearization Method 2 a(443) GSM 6-Band a(443) Global bb(443) Global LM LM SVD SVD IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Matrix Inversion Still Missing Highs in a & bb NOMAD aph(443) Eutrophic Waters, 5-Band GSM LM vs SVD SVD Amoeba 6-Band a(443) 5-Band 4-Band bb(443) 3-Band IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Uncertainties IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Standard Deviation of Rrs DistributionSeaWiFS March 2005 443 412 490 510 555 670 0.005 0.0 reflectance units IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Misc IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Trophic Subsets Deep-Water (Depth > 1000m) Oligotrophic (Chlorophyll < 0.1) Eutrophic (1 < Chlorophyll < 10) Mesotrophic (0.1 < Chlorophyll < 1) IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Salinity IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
ap* I used the Bricaud function to compute aph* for 25 chl concentrations between 0.05 and 3 (evenly distributed in log space), then computed the average spectra and spit out the 10nm wide aph* values for SeaWiFS wavelengths: In the attached plot, the average aph* spectra is the white line, the 10nm (wvl-5 <= wvl < wvl+5) version is in blue, and GSM is in red. IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
Model Differences: Global View IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
QAA vs 6-Band GSM: a(443) & a(555) Eutrophic Mesotrophic Oligotrophic GSM 443 QAA 555 IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG 1.05 aw
QAA - GSM: a(443) & a(555) Eutrophic Mesotrophic Oligotrophic 443 555 IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
QAA vs 6-Band GSM: bb(443) & bb(555) Eutrophic Mesotrophic Oligotrophic GSM GSM 443 QAA QAA 555 IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG
QAA - GSM: bb(443) & bb(555) Eutrophic Mesotrophic Oligotrophic 443 555 IOP Algorithm Workshop, Ocean Optics XIX, 3-4 Oct 2008, B. Franz, NASA/OBPG