Multi-Angle Implementation of Atmospheric Correction for MODIS:

Algorithm MAIAC

MODIS Science Team Meeting

Alexei Lyapustin and Yujie Wang, GEST UMBC/NASA GSFC, mail code 614.4

January 26, 2010

Block-diagram of Processing(Backup: Standard Algorithm)

New Granule

1. Grid L1B Data and Split in Tiles

3. QB, LTP: Cloud Mask (Covariance-Based Algorithm)

2. LTP: RetrieveWater Vapor

(NIR Algorithm)

Queue of K days

Is B7 BRF known?

YesUse MODIS Dark Target Algorithm

7. QP: Retrieve BRF and albedo in reflective bands.

Backup: Lambertian retrieval

6. LTP: Retrieve “Angstrom exponent”

5. LTP: Retrieve B3 AOT using known surface BRF,

B=7bB

4. QB: Retrieve Spectral Regression Coefficients in

Blue band B3 (bB).

No

4a. Retrieve Snow sub-pixel Fraction and

Snow Grain Size

QB:Is Snow Detected?

MAIAC Products (1 km, gridded)Atmosphere: Surface:

• Cloud Mask; Parameters of RTLS BRF model;• Water Vapor; Surface Reflectance (BRF)/ Albedo;• AOT & fine mode fraction; Dynamic Land-Water-Snow Mask.

• Basis - covariance analysis (identifies reproducible surface pattern in the time series) & reference clear-sky image of surface (B. Rossow)

• High covariance - CLEAR. Ephemeral clouds disturb the pattern and reduce covariance.

• Because covariance removes the average component of signal and uses variation, it works well: 1) for bright surfaces and snow, 2) in high AOT conditions if the surface variability is still detectable.

• Algorithm maintains a dynamic clear-skies reference surface image (refcm), used as a comparison target in cloud masking.

• Internal Land-Water-Snow surface classification.

CM Legend:Blue (Clear), Red & Yellow (Cloudy). DOY 187

GSFC, USA

DOY 116Mongu, Zambia

DOY 139Solar Village

Cloud MaskLeft – MODIS TERRA RGB, 2003 (5050km2), Right – MAIAC CM

(BTij < BTG – 4) AND (r1ij >refcm.r1ij+0.05) PCLOUD Bright-Cold Cloud Test (@Pixel)

Approach:

- Accumulate gridded MODIS L1B data for K days;

- Process K days for area NN pixels simultaneously:

KN2 (measurements) > K + 3N2 (unknowns), if K>3- Derive shape of BRF from 2.1 m, and use spectral scaling to reduce DIM:

KN2 > K + N2

MISR heritage: MODIS heritage:-Using spatial and angular structure of imagery for aerosol - RTLS BRF retrieval algorithm (Schaaf et al., 2002)

retrievals (Martonchik et al., IEEE TGARS 1998); - Gridding algorithm (Wolfe et al., 1998)

-Using angular and spectral shape similarity constraints - Cloud Mask (Ackerman et al., 1998)

in aerosol retrievals over land (Diner et al., RSE, 2004).

Basis:- surface is spatially variable and stable in short time

intervals;- aerosols are variable in time and have a mesoscale

(60-100 km) range of global variability.

7Bijij

Blueij b

Generic Retrieval of SRCQueue of K

days

1

2

3

kN2

}),,({ ijijvgoLij bkkk

{k} {bij}

Prescribed vs Retrieved SRC(Example for Goddard Space Flight Center, 2000, DOY 84-93)

Initial MAIAC run Second run after initialization

TOACM

C5 AOT TOACM AOT

NBRF SRCBRF

1. In standard retrievals, AOT correlates with surface brightness (left).

2. MAIAC removes artificial correlation by means of SRC retrieval (right).

Aerosol Retrieval Algorithm• Compute AOTB and weight of coarse mode using Blue (B3), Red (B1),

SWIR (B7) bands.

• Surface BRF: use SRC in blue band, . BRF in B1 and B7

is known from previous retrieval with uncertainty at TOA.

• Algorithm: Fit Blue band to find AOTB for given , and find by minimizing 2

),(

B

TheorMeas AOTRRrmse

7Bijij

Blueij b

TO

A R

efl

ect

an

ce

II

, mBlue Red SWIR

=0, Fine mode

=0.2, Low % of Coarse mode

=0.8, High % of Coarse mode

Water Cloud model

)( ij

… examplesIllustration of AOT and coarse mode fraction

for GSFC in June 2002:

1. Fine mode aerosol does not affect B7 (measured and modeled reflectance agree).

2. Coarse mode aerosol affects B7 (measured reflectance is higher than modeled reflectance)

Resolving thin clouds using Cloud Model(yellow color)

TOA CM NBRFBRFSRC AOT TOA7RTLS7

1

1

1

1

2

2

2

classification: Fine – Coarse - Cloud

• Compute 3 parameters of Ross-Thick Li-Sparse (RTLS) model by fitting

4-16 days of MODIS data at TOA:

Quality Control

- Detect seasonal and rapid surface change from measurements (green-up or senescence shows as large-scale correlated changes in NDVI and NIR and SWIR reflectance as compared to theoretical RTLS values). If surface is stable, use 16-day Queue. If change is detected, use last 4 days for faster response.

- In stable conditions, require consistency with previous solution.

- Check shape of BRDF, rmse etc. (including sufficient angular sampling,

filtering of high AOT …).

Surface Retrieval Algorithm

),(),,(),,(),(),,(),,( 000000 nlGGVVLLD RFkFkFkRR

DOY 233-248 DOY 249-270, 2000

Response to Rapid Surface Change

TOA RGB NBRFBRF

SRCAOT

TOA RGB NBRFBRF

SRCAOT

Example: MODIS PRI analysisHilker, T. et al. (2009).  An assessment of photosynthetic light use efficiency from space: Modeling the atmospheric and directional impacts on PRI reflectance. RSE, doi:10.1016/j.rse.2009.07.012.

Photochemical Reflectance Index (PRI):

The “ocean fluorescence” band 531 nm is sensitive todown-regulation of plant photosynthesis changing by several tenths of 1%, while reference band is stable. The ground measurements showed a good correlation of PRI with light use efficiency ().

554531

554531

Ground PRI (AMSPEC) vs MODIS PRI generated with 6S and MAIAC(Vancouver Island, BC, Canada, March-October 2006)

6S, all angles 6S, backscattering angles MAIAC

Problem of Brighter Surfaces

0

0.2

0.4

0.6

0.8

1

-60 -40 -20 0 20 40 60VZA

Sp

ec

tra

l R

ati

o

NBRF_Red = 0.36NBRF_Blue = 0.16NBRF_B7 = 0.52

135o45o

SZA=30o

SZA=60o

• MAIAC overestimates AOT in the backscattering directions, and underestimates it at forward scattering angles.

• Reason: spectral invariance assumption

is not very accurate for soils and sands when difference in brightness is significant.

• Analysis shows that

the Blue band BRF is more anisotropic than at 2.1 m. The bottom Figure shows Spectral Ratio for the blue (B3/B7) and red (B1/B7). The BRF was computed using LSRT parameters of ASRVN.

7Bijijij b

Desert Pixel

Wang, Y., A. Lyapustin, J. L. Privette, J. T. Morisette, B. Holben, Atmospheric Correction at AERONET Locations: A New Science and Validation Data Set, . IEEE Trans. Geosci. Remote Sens., in press.

AERONET

MA

IAC

Assessment of Lambertian Biasesfrom ASRVN Data (GSFC site)

y = 0.939x + 0.0028

R2 = 0.9936

0

0.04

0.08

0.12

0.16

0 0.04 0.08 0.12 0.16

ASRVN IBRF

RedAOT<0.3SZA<45VZA<45

y = 0.8992x + 0.0058

R2 = 0.9853

0

0.05

0.1

0.15

0.2

0 0.05 0.1 0.15 0.2

ASRVN IBRF

RedAOT<0.3SZA>45VZA>45

y = 0.8514x + 0.007

R2 = 0.9638

0

0.04

0.08

0.12

0.16

0 0.04 0.08 0.12 0.16

ASRVN IBRF

RedAOT>0.3SZA<45VZA<45

y = 0.8306x + 0.0089

R2 = 0.9517

0

0.05

0.1

0.15

0.2

0 0.05 0.1 0.15 0.2

ASRVN BRF

RedAOT>0.3SZA>45VZA>45

y = 0.848x + 0.0085

R2 = 0.9812

0

0.04

0.08

0.12

0.16

0 0.04 0.08 0.12 0.16

ASRVN IBRF

GreenAOT<0.3SZA<45VZA<45

y = 0.7683x + 0.014

R2 = 0.9639

0

0.05

0.1

0.15

0.2

0 0.05 0.1 0.15 0.2

ASRVN IBRF

GreenAOT<0.3SZA>45VZA>45

y = 0.6807x + 0.0191

R2 = 0.9127

0

0.04

0.08

0.12

0.16

0 0.04 0.08 0.12 0.16

ASRVN IBRF

GreenAOT>0.3SZA<45VZA<45

y = 0.6383x + 0.0219

R2 = 0.9121

0

0.05

0.1

0.15

0.2

0 0.05 0.1 0.15 0.2

ASRVN BRF

GreenAOT>0.3SZA>45VZA>45

ASRVN BRF

AS

RV

N L

ambe

rt R

efle

ctan

ce

Slope of regression decreases by up to ~15% in red and 35% in greenfrom consistent comparison of ASRVN-based BRF and RL (Wang, Lyapustin, Privette, Vermote, Schaaf, Wolfe, R. Cook et al.)

y = 0.9734x + 0.0071

R2 = 0.9932

0.1

0.2

0.3

0.4

0.5

0.1 0.2 0.3 0.4 0.5

ASRVN IBRF

MO

D0

9 R

efle

cta

nce

(D

aily

)

NIR

y = 0.9119x + 0.005

R2 = 0.9572

0

0.05

0.1

0.15

0.2

0 0.05 0.1 0.15 0.2

ASRVN IBRF

MO

D0

9 R

efle

cta

nce

(D

aily

)

Red

y = 0.6212x + 0.0279

R2 = 0.8727

0

0.05

0.1

0.15

0.2

0 0.05 0.1 0.15 0.2

ASRVN BRF

MO

D0

9 R

efle

cta

nce

(D

aily

)

Green

y = 1.0061x + 0.0058

R2 = 0.985

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

ASRVN NDVI

MO

D0

9 N

DV

I (D

aily

)

NDVI

Assessment of Lambertian Biasesfrom MODIS Data

ASRVN BRF ASRVN NDVI

Konza EDC

y = 0.9374x + 0.0112

R2 = 0.9926

0.1

0.2

0.3

0.4

0.5

0.1 0.2 0.3 0.4 0.5

ASRVN IBRF

MO

D0

9 R

efle

cta

nce

(D

aily

)

NIR

y = 0.8653x + 0.0039

R2 = 0.94

0

0.03

0.06

0.09

0.12

0.15

0 0.03 0.06 0.09 0.12 0.15

ASRVN IBRF

MO

D0

9 R

efle

cta

nce

(D

aily

)

Red

y = 0.5903x + 0.0187

R2 = 0.8426

0

0.03

0.06

0.09

0.12

0 0.03 0.06 0.09 0.12

ASRVN BRF

MO

D0

9 R

efle

cta

nce

(D

aily

)

Green

y = 0.974x + 0.0284

R2 = 0.9792

0.2

0.4

0.6

0.8

1

0.2 0.4 0.6 0.8 1

ASRVN NDVI

MO

D0

9 N

DV

I (D

aily

)

NDVI

Walker Branch

ASRVN BRFASRVN BRF

Observed biases over vegetated sites are similar to results of ASRVN-based comparison of BRF and RL which indicates that they are mostly due to Lambertian assumption.

AERONET Validation, 7+ yrs. (old Version)