inter-comparison of atmospheric correction over moderately to very turbid waters cédric jamet...

Inter-comparison of atmospheric correction over

moderately to very turbid watersCédric Jamet

Kick-off meetingWimereux, France

May, 14, 2014

Rationale• 15 years of continuous ocean satellite

data:

– ~13 years for SeaWiFS– ~10 years for MERIS– ~10 years of MODIS-AQUA (and still counting)– VIIRS (2012-)– OLCI (2015-)– GOCI (2010- )

Possibility to study inter-annual and decennal variability of biogeochemical parameters in open but also in coastal waters

RationaleNeed for accurate atmospheric correction algorithms

Need to look at the existing algorithms

• Several AC developed for the past 12 years no assessment of differences

• Future ocean color sensors: high spatial resolution, high radiometric resolution, other wavelengths

Highest possibility to study coastal waters

• Need for guidances on using the already developed AC (see number of requests on the forum of oceancolor.gsfc.nasa.gov)

Courtesy of PJ Werdell

no one uses the “black pixel assumption” anymore

SeaWIFS Atmospheric Correction Algorithms

• Three NIR ocean contribution removing/AC algorithms (Jamet et al., RSE, 2011)

– Stumpf et al. (2003)/ Bailey et al., (2010) S03: • Based on Gordon and Wang atmospheric correction (GW94)• SeaWiFS/MODIS standard algorithm• Iterative process• Bio-optical model used to determine bb(670)

– Ruddick et al. (2000) R00:• Based on Gordon and Wang atmospheric correction (GW94)• Spatial homogeneity of the Lw(NIR) and LA(NIR) ratios over the

subscene of interest : Ratio of Lw(NIR) cst = 1.72• ε: Ratio of LA(NIR) determined for each subscene

– Kuchinke et al. (2009) K09:• Spectral optimization algorithm• Junge aerosol models• GSM bio-optical model (Garver, 2002)• Atmosphere and ocean coupled

ρ'w(λ

VIS)

ρw(λ

red)

Chla

Black Pixel Assumption ρw(λ

NIR) = 0

Chla based bio-optical model

First guess in ρw(λ)

ρw(λ

NIR)

ρa(λ)+ρ

ra(λ)=ρ

rc(λ)-ρ

w(λ)

ρrc=ρ

t-ρ

r-tρ

wc-Tρ

g

Remove white caps, sun glint, ozone and Rayleigh scattering

Modelled ρw(λ

NIR)

ρw(λ)Second guess in ρ

w(λ)

Gordon and Gordon and Wang, 1994Wang, 1994

NIR modelling NIR modelling schemescheme

ρa(λ)+ρ

ra(λ)=ρ

rc(λ)

Black Pixel Assumption ρw(λ

NIR) = 0

ε(λNIR1

,λNIR2

)=

[ρa(λ

NIR1)+ρ

ra(λ

NIR1)]/[ρ

a(λ

NIR2)+ρ

ra(λ

NIR2)]

ρa(λ)+ρ

ra(λ)=ρ

rc(λ)-ρ

w(λ)

ρw(λ

NIR)

ρrc=ρ

t-ρ

r-tρ

wc-Tρ

g


ρw(λ)

First guess in ρa(λ)+ρ

ra(λ)

Gordon Gordon and and

Wang, Wang, 19941994

ρ'a(λ)+ρ'

ra(λ)=ρ

rc(λ)

a =ρw(λ

NIR1)/ρ

w(λ

NIR2)

Spatial homogeneity in aerosolsNIR similarity spectrum

NIR modelling NIR modelling schemescheme

DATA• Satellite data:

– (M)LAC SeaWiFS 1km at nadir Processed with SeaDAS 6.1 (R2009) (Fu et al., 1998)

– nLw(412865), (865), (510,865)

• In situ data: AERONET-OC network (Zibordi et al., 2006, 2009)

– Three sites – 7 λ centered at 412, 443, 531, 551, 667, 870

and 1020 nm

Results• Only turbid waters (Robinson et al., 2003):

nLw(670)>0.186• Comparison of the normalized water-leaving radiances

nLw between 412 and 670 nm and of the aerosol optical properties (the Ångström coefficient (510,865) and the optical thickness (865))

MVCO AAOT COVE TOTAL nLw(412) <0

S03 20 163 18 201 7

R00 19 129 17 165 6

Kuchinke 13 134 13 160 0

# of matchups for each algorithm and each AERONET-OC site

Jamet et al., 2011

Conclusions (1/3)

• Comparison of 3 SeaWiFS Atmospheric Correction algorithms

– SeaWiFS standard algorithm: best overall estimates

– Ruddick algorithm: less accurate– Kuchinke: good estimates for short

wavelengths

Conclusions (2/3)• Sensitivity tests on assumptions of each

algorithm

– Ruddick: High impact of a bias of the value of the aerosol ratio ε

– Stumpf R2009: Low impact of the new aerosol models for nLw; Moderate to High impact of the new definition of bb(670)

– Kuchinke: High impact of the Junge aerosol models; “high impact” of the bio-optical parameters in GSM need to tune the bio-optical models

Conclusions (3/3)

• Time processing:

– Standard algorithm: fastest algorithm– Kuchinke: very time consuming (as any

optimization technique)– Ruddick: twice slower than standard

algorithm (need to process two times the same image)

MODIS Atmospheric Correction Algorithms

• Four NIR ocean contribution removing/AC algorithms (Goyens et al., RSE, 2013)– Stumpf et al. (2003)/ Bailey et al., (2010) S03:

• Based on Gordon and Wang atmospheric correction (GW94)• SeaWiFS/MODIS standard algorithm• Iterative process• Bio-optical model used to determine bb(670)



– Wang and Shi (2005, 2009):• Based on GW94• SWIR bands for determination of aerosol models: black pixel

assumption– Shroeder et al. (2007)

• NN direction inversion• Coupled ocean-atmosphere

NIR/SWIRSwitching algorithms based on

Turbid Water Index

Tind

~ ρrc(748, 1240)

GW94-based algorithm

assuming ρw(λ

SWIR)=0

ρrc=ρ

t-ρ

r-tρ

wc-Tρ

g


ρw(λ)

STD AC method

SWIR AC method

~ Gordon ~ Gordon and and

Wang, Wang, 1994 1994

MODIS Atmospheric Correction Algorithms

• Four NIR ocean contribution removing/AC algorithms (Goyens et al., RSE, 2013)– Stumpf et al. (2003)/ Bailey et al., (2010) S03:

• Based on Gordon and Wang atmospheric correction (GW94)• SeaWiFS/MODIS standard algorithm• Iterative process• Bio-optical model used to determine bb(670)



– Wang and Shi (2005, 2009):• Based on GW94• SWIR bands for determination of aerosol models: black pixel

assumption– Shroeder et al. (2007)

• NN direction inversion• Coupled ocean-atmosphere

MVCO

Approach OverviewApproach Overview

c. c. In situ In situ datadata

COVE

HELSINKI

• AERONET-OC data

MVCO


c. c. In situ In situ datadata

COVE

HELSINKI

• AERONET-OC data Moderately turbidModerately turbid


c. c. In situ In situ datadata• In situ MUMM & LOG data

TriOS RAMSES above and in water measurementsTriOS RAMSES above and in water measurementsCOVE

ModeratelyModerately to to eextremely turbidxtremely turbid



20 spectra 76 spectra

21 spectra 93 spectra

Hyperspectral TriOS ρHyperspectral TriOS ρww(λ) LOG data(λ) LOG data



Hyperspectral TriOS ρHyperspectral TriOS ρww(λ) MUMM data(λ) MUMM data

131 spectra with moderately to extremely turbid waters



French Guiana 2012

RV Melville – South Atlantic 2011

I. I. EvaluationEvaluation and comparisoncomparison of turbid water AC methods

a. Global Evaluationa. Global Evaluation

Goyens et al., (2013) [a]

MUMM 211 AERONET-OC & LOG

match-ups Overall the STD AC method

performs the best ↑ spatial coverage but ↑ neg.

Lwn

(λ) values with the MUMM AC

method STD, MUMM and NIR-SWIR tend

to underestimate Lwn

(λ) (bias

between -26 and 3%) STD, MUMM and NIR-SWIR

perform better in the green (< 20% RE), not as well in the blue & red (> 30% RE)

NN performs better in the blue (30% RE) and red (22% RE) and not as well in the green (up to 27% RE)


a. Global Evaluationa. Global EvaluationMUMM

211 match-ups AERONET-OC & LOG)

Overall the STD AC method performs the best

↑ spatial coverage but ↑ neg. L

wn(λ) values with the MUMM AC



(λ) (bias



NN performs better in the blue (30% RE) and red (22% RE) and not as well in the green (up to 27% RE) Goyens et al., (2013) [a]






wn(λ) values with the MUMM

AC method STD, MUMM and NIR-SWIR tend


(λ) (bias










method

STD, MUMM and NIR-SWIR tend to underestimate L

wn(λ)

(bias between -26 and 3%) STD, MUMM and NIR-SWIR











(λ) (bias








↑ spatial coverage but ↑ neg. Lwn(λ) values with the MUMM AC


to underestimate Lwn(λ) (bias




Goyens et al., 2013


b. b. Water class-Water class-specific validationspecific validation

Goyens et al., (2013) [a]Normalized LOG Rrs(λ) data spectra per class according to the classification scheme of Vantrepotte et al. (2012)

Detrital/mineral particles & low phytoplankton

Phytoplankton

CDOM


Goyens et al., 2013; Vantrepotte et al., 2012

(CDOM & phyto)(detrital & mineral) (phyto)




Detrital/mineral particles & low phytoplankton

- NN AC method shows the best performances (RE from 16% to 27% and bias from -7% to 5%)

- Among the GW94-based AC methods, the MUMM approach results in the best performances (RE from 20% to 43% and bias from -42% to -19%)





- Overall highest RE (up to 73%) and bias (up to -26%) are encountered over CDOM dominated waters

- STD AC method shows the best performances (RE from 14% to 53% and bias from -16% to 14%)

CDOM


- Overall the STD AC method performs the best

- The MUMM AC method ensures a large spatial coverage (but retrieves more neg. ρw(λ) values)

- For water masses optically dominated by detrital/mineral particles the MUMM AC method performs better

In situ-satellite match-up pairs

In situ ρw(λ) vs.

satellite ρw(λ)

Global validation

Water class-specific validation

I. Evaluation and comparison of existing turbid water AC methods

Determine best Determine best AC method(s) AC method(s) to improve !to improve !


c. c. ConclusionConclusion

Conclusions• Several methods exist today Difficult to estimate

which one is the best

• All methods have advantages and limitations

• Not sure accuracy is reached for short visible wavelengths

• Still work to do or the actual accuracy is enough for bio-optical applications???

My participation• Chairman:

– Management of the WG– Editor of the report– Organizer of annual and teleconf meetings

• Processing datasets with Goyens et al. (2014) AC• Collection of in-situ datasets

• Full analysis of the outputs of the AC

– Match-ups– Transects– Sensitivities studies– Time series

Your participation• Developers provide their algorithm (+ process datasets)

• How do you want to participate?– K. Stamnes: algorithm + RT model for atmosphere– X. He: algorithm + RT model for atmosphere– S. Sterckx: adjacency effects comparison + extreme turbid

waters in-situ dataset– P. Shanmugam: algorithm + in-situ dataset in Indian Ocean– J. Brajard: algorithm– S. Bailey: NASA in-situ datasets + implementation of AC in

SeaDAS + algorithm– K. Ruddick: link with IOCCG WG + in-situ datasets– T. Schroeder: algorithm + in-situ datasets around Australia– M. Wang: experience from report #10 + algorithms

Why a new IOCCG WG?

• Complement of the IOCCG report #10: « Atmospheric Correction for Remotely-Sensed Ocean-Colour Products » (Wang, 2010)Update (could also be an update of IOCCG

report #3)

• This WG focused mainly on open ocean waters• And in coastal waters??• Need for guidances on using the already

developed AC

Purpose of this new WG

Summary• Goal: Inter-comparison and evaluation of existing

AC algorithms over turbid/coastal waters Understanding retrievals differences between

algorithms• Challenge: to understand the advantages and

limitations of each algorithm and their performance under certain atmospheric and oceanic conditions

• Only focus on AC algorithms that deal with a non-zero NIR water-leaving radiances.

• High demand for AC guidelines• Outputs timely Guidances on the use of AC over turbid waters Recommendations for improving and selecting

the optimal AC• Not a sensor-oriented exercise

Questions• How well do the aerosols need to be retrieved?• Does the technique matter?• How good is the extrapolation from SWIR to

NIR to VIS?• How to improve AC with the historic

wavelengths?– Better bio-optical models– New contrains (such as relationships between Rrs,

cf Poster of Goyens et al.)– Regional AC?

• Do we really need extra wavelengths in NIR/SWIR? What info can offer the future satellite sensors?

Evaluation of AC• Can round robin lead to improvements?

– What can we learn?• Range of validity and advantages• Limitations

– Sensitivity studies• Fixed aerosols Variation/change of the bio-optical

model• Fixed bio-optical model Variation/change of the

aerosol models – Uncertainties propagation and budget on the hypothesis

• Ruddick et al. (2000)• Bayseian statistics for NN (Aires et al., 2004a, 2004b,

2004c)– Uncertainties on the NN parameters (weights)– Uncertainties on the outputs

• Others ?

Evaluation of AC

• Can round robin lead to improvements?– What can we learn?

• Drawbacks and advantages• Limitations

– Sensitivity studies• Fixed aerosols Variation/change of the bio-optical

model• Fixed bio-optical model Variation/change of the

aerosol models – Uncertainties propagation and budget on the hypothesis

• Ruddick et al. (2000)• Neukermans et al., (2012)• Bayseian statistics for NN (Aires et al., 2004a, 2004b,

2004c)– Uncertainties on the NN parameters (weights)– Uncertainties on the outputs

Classification of Lw spectra per water type

- 4 water type classes defined by Vantrepotte et al. (2012)

- Novelty detection technique:Assigns each spectra to one

of the 4 water type classes

using the Mahalanobis distance

METHODS

Focus on turbid waters only !Distinguish classes based on

normalized reflectance spectra

D'Alimonte et al. (2003)

Evaluation of the algorithms as a function of the water type

RESULTS

We don't have any mathup for Class 3 (very turbid water masses)!

Goyens et al., 2013

Rationale• Coastal waters more and more investigated in ocean color

• But more challenging than open ocean waters:

– temporal & spatial variability• satellite sensor resolution• satellite repeat frequency• validity of ancillary data (SST, wind)• resolution requirements

– Land contamination (adjacency effects)– non-maritime aerosols (dust, pollution)

• region-specific models required?• absorbing aerosols

– suspended sediments & CDOM• complicates estimation of Rrs(NIR)

• complicates BRDF (f/Q) corrections• sensor saturation

– shallow waters

Need for accurate data processing algorithms








– shallow waters


Robert Frouin








– shallow waters


Sean Bailey






– suspended sediments & CDOM

•complicates estimation of Rrs(NIR)• complicates BRDF (f/Q) corrections• sensor saturation

– shallow waters


ρt: total (measured)

ρr: Rayleigh (known)

ρa: aerosols (unknown)

ρra: rayleigh-aerosols (unknown)

ρwc:whitecaps (modelled/wind)

ρw: water (unknown, 10% of ρt)

ρg: glitter (masked or modelled

Objectives: 5% for absolute radiances and 2% for relative reflectances

Rationale of atmospheric correction

ρt: total (measured)

ρr: Rayleigh (known)

ρa: aerosols (unknown)

ρra: rayleigh-aerosols (unknown)

ρwc:whitecaps (modelled/wind)

ρw: water (unknown, 10% of ρt)

ρg: glitter (masked or modelled)

Objectives: 5% for absolute radiances and 2% for relative reflectances

Rationale of atmospheric correction

Classic Black Pixel Assumption (1/2)NIR bands

670nm 765nm 865nm

Hypothesis: absorbing ocean w negligeable

t()-r()=ra(a(A()

Classic Black Pixel Assumption (1/2)NIR bands

670nm 765nm 865nm


t()-r()=ra(a(A()

calculate aerosol ratios, :

(748,869) ~

(,869) ~

a(869)

a(748)

a(869)

a()

Classic Black Pixel Assumption (1/2)

Determination of aerosols

models

Calculation of ρA and t

NIR bands

670nm 765nm 865nm


t()-r()=ra(a(A()

Classic Black Pixel Assumption (2/2)

412nm 443nm 490nm 510nm 555nm

Visible bands, t()-r()-A())/t()= w()

Courtesy of PJ Werdell

no one uses the “black pixel assumption” anymore

many approaches exist, here are a few examples:• assign aerosols () and/or water contributions (Rrs(NIR)) e.g., Hu et al. 2000, Ruddick et al. 2000• use shortwave infrared bands e.g., Wang & Shi 2007• use of UV bands He et al., 2012

• correct/model the non-negligible Rrs(NIR)

Lavender et al. 2005 MERIS Bailey et al. 2010 used in SeaWiFS Reprocessing 2010 Moore et al. 1999 MERIS/OLCI Wang et al. 2012 GOCI• use a coupled ocean-atmosphere optimization/inversion e.g., Chomko & Gordon 2001; Stamnes et al. 2003; Jamet et al., 2005; Ahn and Shanmugam, 2007; Doerffer & Schiller, 2008; Kuchinke et al. 2009; Schroeder et al., 2007; Steinmetz et al., 2010; Brajard et al., 2012

Approaches to account for Rrs(NIR) > 0 sr-1 overlap

Courtesy of Jeremy Werdell

Evaluation• Evaluation of AC already exists in

journals:– SeaWiFS: Zibordi et al. (2006, 2009); Banzon et al.,

2009; Jamet et al. (2011); – MODIS-AQUA: Zibordi et al. (2006, 2009); Wang et

al. (2009), ;Werdell et al., 2010; Goyens et al. (2013)– MERIS: Zibordi et al. (2006, 2009); Cui et al. (2010);

Kratzer et al. (2010); Melin et al. (2011)

• BUT most of the time only about one specific AC (with eventually comparisons with the official AC)

• Only few papers on round-robin



Lavender et al. 2005 MERIS Bailey et al. 2010 used in SeaWiFS Reprocessing 2010 Moore & Lavender, 1999 MERIS/OLCI Wang et al. 2012 GOCI• use a coupled ocean-atmosphere optimization/inversion e.g., Chomko & Gordon 2001; Stamnes et al. 2003; Jamet et al., 2005, Ahn and Shanmugam, 2007; Doerffer & Schiller, 2008; Kuchinke et al. 2009; Schroeder et al., 2007; Steinmetz et al., 2010; Brajard et al., 2012





Lavender et al. 2005 MERIS Bailey et al. 2010 used in SeaWiFS Reprocessing 2010 Moore & Lavender 1999, MERIS/OLCI Wang et al. 2012 GOCI• use a coupled ocean-atmosphere optimization/inversion e.g., Chomko & Gordon 2001; Stamnes et al. 2003; Jamet et al., 2005; Ahn and Shanmugam, 2007; Doerrfer & Schiller, 2008; Kuchinke et al. 2009; Schroeder et al., 2007; Steinmetz et al., 2010; Brajard et al., 2012



Kratzer et al., 2010

He et al., 2012

USE OF UV BANDS

inter-comparison of atmospheric correction over moderately to very turbid waters cédric jamet...

Documents