Assessing  MODIS  C06  Urban  Correc6ons    Using  the  High  Resolu6on  Dragon  AERONET  Network  

 Nabin  Malakar,  Adam  A/a,  Barry  Gross,  Fred  Moshary    

Op#cal  Remote  Sensing  Lab,  CCNY  Min  Oo  

 CIMSS  /  UW-­‐Madison    

Mo6va6on  l Aerosol  Retrieval  over  land  is  greatly  affected  by  land  surface  albedo  (if  bright  enough).  

l MODIS  land  surface  compensa6on  algorithms  for  global  applica6ons  were  trained  using  non  urban  land  surface    types  (mixtures)  such  as  vegeta6ons/  clays.  

l As  urbaniza6on  con6nues  to  increase,  the  differences  in  land  surface    behavior  need  to  be  bePer  understood.    

l These  issues  become  even  more  significant  as  higher  resolu6on  aerosol  products  such  as  C006  3km  Aerosol  Retrievals  become  available  

Single  scaPering  Mul6ple  ScaPering      

Photons  hit  land  surface  And  reflected  back  to  space  

AOD  Bias  (Dragon  Network)    

3km  product   10km  product  

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

AERONET

MO

DIS

C006

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

AERONET

MO

DIS

C005

Clear  Biases  seen  in  the  products  but  enhanced  at  3km    

Approach  

l  We  previously  inves6gated  the  existence  of    high  bias    in  AOD  retrievals  in  C005  for  significantly  urbanized  areas  such  as  New  York  City  

l  By  combining  AERONET  with  MODIS  observa6ons  over  sufficiently  “clean”  days,  it  is  possible  to  improve  on  the  exis6ng  land  surface  model  needed  to  correct  for  land  reflec6on  

l  Applying  this  approach  over  a  region  is  complicated  by  the  fact  that  only  a  single  AERONET  sta6on  is  available  and  an  assump6on  that  the  AOD  and  phase  func6on  proper6es  are  homogeneous  on  a  regional  scale  are  clearly  an  issue  

l  Using  the  Dragon  Network  allows  for  the  poten6al  of  using  bePer  AERONET  informa6on  in  “tuning”  the  surface  while  also  providing  bePer  sta6s6cal  valida6on.          

l  We  also  inves6gate  the  neural  network  approach  to  correct  the  bias.  

Opera6onal  satellites  retrieval  over  land  l  MODIS  aerosol  retrieval  uses  three  wavelength  channels  (470,  660,  2120nm)  

l  Mul6  wavelength  measurements  help  separate  fine  /  coarse  components.    

l  However,  the  surface  reflec6on  contaminates  the  signals.  

l  To  es6mate  this,  MODIS  does  the  following    l  Assumes  the  long  wavelength  channel  is  insensi6ve  to  the  atmosphere  so  the  signal  must  be  due  only  to  the  ground  reflec6on  (Rg_2120)  

l  Once  the  long  wavelength  reflec6on  is  es6mated,  use  semi-­‐empirical  models  taking  into  account  how  vegeta6ve  the  surface  is  to  es6mate  the  VIS  to  SWIR  ra6os  (Rg_470)/  (Rg_2120),  (Rg_660)/  (Rg_2120)  

l  MODIS  uses  an  index  called  the  Modified  Vegeta6on  Index  (MVI),  which  combines  NIR  and  SWIR  to  es6mate  vegeta6on  class.    

l  We  demonstrate  that  these  ra6os  are  not  well  represented  in  opera6onal  algorithms  and  need  refinement  which  allows  bePer  aerosol  retrieval.      

TOAm

TOAm

TOAm

TOAmMVI

µµ

µµ

ρρ

ρρ

12.224.1

12.224.1

+

−=

Retrieving  Land  Surface  Band  Spectral  Ra6os  

l  The  Collect  5/6  approach  allows  the  VIS-­‐SWIR  ground  albedo  correla6on  coefficients  to  be  a  func6on  of  surface  type  (urban/vegeta6on  MVI)  and  observa6on  angles  (scaPering  angle).    

l  In  our  case,  we  ingest  AOD  from  Aeronet  to  atmospherically  correct  the  MODIS  images      

l  To  ensure  that  the  best  surface  retrieval  is  made,  the  following  filters  are  applied    –  AOD  <  0.2,  –  angstrom  exponent    >  1    to  assure  minimal  aerosol  contamina6on  at  2.1  um    

–  Homogeneous  condi6ons  (variability  of  AERONET  AOD  for  +/-­‐  3  hours    <  20%)  which  helps  us  extrapolate  AOD  over  en6re  domain  

–  Mask  all  water  pixels  

–  For  Dragon  Network,  we  use  Aeronet  averages  when  possible  to  improve  quality  of  land  surface  reflec6on  and  remove  homogeneity  assump6on.      

Obtaining  surface  albedos  using  combined    MODIS  –  Aeronet  Data    

( )( )

( ) albedo spherical ic Atmospherontransmissi total downward and Upward ,

ereflectancpath,,,

,

=

=

=Δ

λ

θλ

φθθλρ

sT ud

ivatm

g

udgatmTOA s

TTρ

ρρρ

−+=1

Aeronet  Op6cal  Depth  +  MODIS  Aerosol  Phase  Func6on  consistent  with  AOD  

Once  this  is  done,  we  can  Isolate  Lamber6an  albedo    

)(

atmTOAud

atmTOAg sTT ρρ

ρρρ

−+−

=⇒

Use  Aeronet  AOD  to  fix  the  MODIS  Aerosol  Phase  func6on  model  From  this,  we  can  get  all  relevant  atmospheric  scaPering  parameters    

[⌧550aer

]aeronet

! [Paer

(⇥scat

, ⌧550aer

,�)]urban-nonabs

80 90 100 110 120 130 140 150 1600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Rh

o 0

.66

um

/ R

ho

2.1

2 u

mScattering angle

y = 1.2e-005*x + 0.77

data 1 linear

80 90 100 110 120 130 140 150 1600

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Rh

o 0

.47

um

/ R

ho

2.1

2 u

m

Scattering angle

y = 0.00059*x + 0.45

data 1 linear

Band  Correla6on  with  ScaPering  Angle  

( ) ( ) ( )2:1

2120

=

Θ=

if gsiig ρλρ

 

Mean=0.5153  std=0.0858  

Mean=  0.7734  std=0.0729  

Rho  0.470/  Rho2.12                              Rho  0.660/  Rho  2.120      

Once  new  correla6ons  are  found,  we  can  replace  the  COO5    Correla6on  procedures  and  assess  retrieval  of  AOD  (for  all  cases)    

Band  Correla6on  with  ScaPering  Angle  (water  mask  included)  shows  minimal  angular  dependence  valida/ng  lamber/an  

assump/ons    

General  Rela6onship  between  Surface  Type  and  the  VIS/SWIR  reflec6on  ra6os  in  urban  areas  

           Regional  surface  data  retrievals  (50km  x  50km)    around    different  ci6es  with  AERONET  at  center.    Note  that  VIS/SWIR  ra6os  decrease  with  MVI  index  in  contradic6on  to  the  MODIS  C005  opera6onal  models.  (Later,  we  see  that  C006  trend  is  improved  over  C005)  

• When  MVI  is  low  (i.e  urban),  SRC’s  are  significantly  underes6mated  • The  C005  model  actually  shows  an  opposite  trend  indica6ve  of  the  differences  between  low  MVI  soils  and  urban  materials  • NYC  is  by  far  the  most  biased  region  over  other  urban  areas  in  comparisons  to  other  urban  centers.    

Anomalies  in  Spectral  Ra6os  

Tuned  Surface  Reflec6on  Ra6o  

Strong  correla6on  between  urban  frac6on  and  regionally  tuned  surface  reflec6on  ra6o  

Urban  Land  Cover     Deciduous  broadleaf  forest  

Land  Surface  Spectral  Ra6o  

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

MVI Index

Ref

660

/ R

ef 2

120

croplandmixed forresturban/builtdeciduous broadleaf

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

MVI Index

Ref

660

/ R

ef 2

120

croplandmixed forresturban/builtdeciduous broadleaf

Regional  Surface  Spectral  Ra6o   C006  Surface  Spectral  Ra6o  

Spectral  Ra6os  by  land  class  

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

MVI Index

Ref

660

/ R

ef 2

120

cropland

regionalC006

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

MVI Index

Ref

660

/ R

ef 2

120

mixed forrest

regionalC006

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

MVI Index

Ref

660

/ R

ef 2

120

urban/built

regionalC006

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

MVI Index

Ref

660

/ R

ef 2

120

deciduous broadleaf

regionalC006

C006  generally  does  bePer  than  C005  (correct  trend)    but  urban  land  class  is  completely  underes6mated    at  low  MVI  

Bias  Dependence  on  Different  Factors  

l  Small  but  posi6ve  bias/RMSE  dependence  on  %  urban  and  scaPering  angle  

l  Negligible  bias  on  C006  surface  reflec6on  ra6o  and  angstrom  Coefficient.    

l  Urban  classifica6on  should  be  ingested  into  high  resolu6on  algorithms  

0 20 40 60 80-0.2

0

0.2

0.4

0.6

Urban %

AO

D C

006 -

AE

RO

NE

T A

OD

100 120 140 160 180-0.2

0

0.2

0.4

0.6

Single Scattering Angle%A

OD

C00

6 -

AE

RO

NE

T A

OD

1 1.5 2 2.5-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

Angstrom Coefficient%

AO

D C

006

- A

ER

ON

ET

AO

D

0.4 0.45 0.5 0.55 0.6-0.2

-0.1

0

0.1

0.2

0.3

0.4

660 /2120 Reflectance Ratio

AO

D C

006 -

AE

RO

NE

T A

OD

Case  Scenario    July  29  1740  UTC  

l Strongest  correc6ons  occur  in  urban  zones  

l Best  agreement  seen  when  correc6on  is  applied  

l No  significant  correc6on  in  non  urban  area  (green  circle)    

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

AERONET

MO

DIS

3km

AO

D

July 29 AQUA 1740 UTC

C006Regional

DragonNET  AOD  retrieval  comparison    

-0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.50

10

20

30

40

50

60

70

80

90

100

AOD Bias

Fre

quen

cy

AERONET AOD - C006 AODAERONET AOD - Derived AOD

Significant  improvement  in  BIAS  and  negligible  change  in  correla6on  coefficient  

mean_bias_C006          -­‐0.0815    mean_bias_TunedAOD            -­‐0.0491  

Bias  Correc6on  using  Machine-­‐  Learning  17  

Target  

Compare  Machine-­‐Learning:  

Neural  nets,  SVM,  RF,  GP  etc.  

Input  

Neural  network  18  

yk = �

0

@nX

j=1

wkjxj

1

A

•  Also  referred  to  as  mul6  layer  perceptron  method,    •  Used  widely  for  classifica6on  or  func6on  approxima6on.  

   

Where ø:  is  the  transfer  func6on   wkj: weight from unit j to unit k,��� xj : n input variables

The  output  of  the  kth  neuron:  

Inputs

Hidden layer

Outputs

Tes6ng  Various  Combina6ons  

AOD+Surf_470+Surf_660_Surf2100+ScaPering  AOD+Surf_470+Surf_660_Surf2100  

Tes6ng  Various  Combina6ons  

AOD  +  Lat+Lon+Land  class   AOD  +Surf047_066_213+ScaPAngle+LC  

•  Improved  correla6on  observed  arer  bias  correc6on  •  Correc6on  on  the  overes6ma6on    

Bias  Corrected  AOD  show  good  Correla6on  

Conclusions  l  Assessment  of  3km  resolu6on  products  using  Dragon  Network    shows  somewhat  

enhanced  bias  in  comparison  to  10km    

l  We  find  that  the  regionally  tuned  surface  spectral  ra6o  model  is  highly  correlated  to  several  dis6nguishing  land  classes  (Urban  /  deciduous  broadleaf  forest)  

l  The  current  MVI  parameter  used  to  get  the  VIS  channel  surface  albedo  es6mate  is  qualita6vely  and  quan6ta6vely  insufficient  to  separate  urban  land  areas  from  other  land  classes  (deciduous  broadleaf  forest)    

l  Significant  Improvement  can  be  seen  in  bias  reduc6on    using  regional  land  surface  model  with  negligible  differences  in  correla6on  

l  Adding  land  classifica6on  with  MVI  should  help  remove  anomalies  for  urban  retrievals.    

l  We  used  the  MODIS  3  km  AOD  products  from  AQUA  and  TERRA,  and  developed  a  machine-­‐learning  framework  to  compare  and  correct  the  remote  sensing  product  with  respect  to  the  ground-­‐based  AERONET  observa6ons.  

l  We  also  constructed  a  neural  network  es6mator  to  obtain  bias-­‐corrected  AOD  product.      

Future  Work  l  Es6mate  PM2.5  from  the  bias-­‐corrected  AOD  

l  Par6culates  with  a  diameter  of  2.5  microns  or  less  l  Can  have  adverse  health  effects  l  Once  in  the  body  may  lead  to  oxida6ve  inflamma6on  in  the  organs.  

Ref:  hPp://www.airnow.gov  

Thank  you!