Äîêóìåíò âçÿò èç êýøà ïîèñêîâîé ìàøèíû. Àäðåñ îðèãèíàëüíîãî äîêóìåíòà : http://www.iki.rssi.ru/earth/pres2006/baccini.pdf Äàòà èçìåíåíèÿ: Mon Feb 19 18:36:17 2007 Äàòà èíäåêñèðîâàíèÿ: Tue Oct 2 11:08:13 2012 Êîäèðîâêà: Ïîèñêîâûå ñëîâà: http www.astronomy.ru forum index.php topic 4644.0.html

Overview of MODIS-based mapping of NELDA Land Cover and Approaches to its Validation
Ales sandro Baccin i, Mark Fr ied l, Curtis Woodcock abaccin i@bu.edu Department of Geography Center for Remote Sensing Boston University http ://geo graphy.bu.edu/landco ver/
1


MODIS Regional Land Cover
Objectives :
­ Combine regional expertise and existing products in the mapping procedure ­ Improve and update (circa 2005) of Northern Euroasia land cover characterization ­ Develop a new legend consistent with FAO LCCS
2


What is Land Cover?
· Generalized classification of the biophysical condit ions at the Earth's land surface · Three key d imensions
­ Natural vegetation ­ Barren and unvegetated land areas ­ Developed/Human modified land areas

3


Why Land Cover?
· Global Change Perspective
­ Land con ver s ion and land use by humans represent the largest s in gle mechanism o f en viro nmental chan ge
· · · · · Carbon storage/release Biodiversity La nd resources & food security Hydrolo gy and water resources Etc......
4


Outline
· Introduction and Context · MODIS Land Cover Mapping
­ Description of data sets ­ Classification methods
· Post-Processin g

­ " Validation"

· New Legend LCCS comp liant

5


MOD12Q1: What Is It?
· Land Cover Types
­ IGBP, UMD, LAI/FP AR, BGC, CLM ­ 1 km

· Confidences
­ Class ificatio n confidence (percent scale) for each pixel

· Secondary IGBP Label
­ For IGBP, a secondary class labe l for each pixe l

6


IGBP Land Cover Units (17)
(Primary Layer) · Natural Vegetation (11)
­ Evergreen Needleleaf Forests ­ Evergreen Broadleaf Forests ­ Deciduous Needleleaf Forests ­ Deciduous Broadleaf Forests ­ Mixed Forests ­ Closed Shrublands ­ Open S hrublands ­ Woody Savannas ­ Savannas ­ Grasslands ­ Permane nt Wetlands

· De ve loped and Mosa ic Land s (3)
­ Croplands ­ Urban a nd B uilt-Up La nds ­ Cropland/Natural Vegeta tion Mosaics

· Non ve getated Lands (3)
­ Snow and Ice ­ Barren ­ Water Bodies

7


8


Global Land Cover Classification Methods
Three main components
1. Exploits spectral and temporal information from MODIS 2. Robust, repeatable classification algorithm 3. Requires extensive, high quality training site data base (STEP)

9


Data
· MODIS Data
­ 32-day Normalized BRDF-Adjusted Reflectances (NBARs) asse mbled over one year of observations ­ 7 spectral bands, 0.4­2.1 µm, similar to Landsat ­ 32-day Enhanced Vegetation Index (EVI)

· Training Data
­ 2130 training sites delineated from high resolution satellite imagery (largely Landsat)
10


Inputs and Classification Flow
(Friedl et al. 2002; RSE)

·

Features From MODIS:
­ Temporal and spectral inf ormation ­ 12 (annual) 32-day composites

Ext ract Exe mplars From STEP Database Estimate Classification Apply Classification to Global/Regio nal Data Fuse Results With Ancillary Data (post-processing)

· Surface Reflectance (NBAR)
­ View-angle corrected surf ace ref lectance ­ 7 land bands

· Enhanced Vegetatio n Index (EVI)
­ Computed from NBARs

· Ann ual Metr ics
­ Min, max, mean f or each band

Maps

11


Key Input Used for Classification: NADIR, BRDF-Ad justed Reflectance
(Schaaf et al., 2002; RSE)

Remo ves artifacts assoc iated w ith var iab le view geometry

12


Classification Algorithm
Decision Tree
· C4.5: Univariate Decision Tree · Nonparametric · Boosting

· Provides robust, repeatable results · Relies hea vily on input training database
13


Decision Tree Classification
(Friedl and Brodley, 1997 ; RS E)

· Goa l:
­ Optimal prediction of class labels from a set of feature values

· Bas ic approach
­ Supervised learni ng using traini ng data

· Key attributes :
­ Nonpara metric ­ Able to handle noisy or missing features ­ Adept at capturing non-li near, hierarchical patterns

14


Optimizing Classification: Boosting
(Mc Iver and Friedl, IEEE TGARS 2001)

· Estimate multiple trees
­ At each iteration, reweigh t samp le to focus on difficu lt cases

Basic Algorithm
1. Initia lize w(i)t=1/N 2. At each iteration : 1. t = w(i) for incorrect predictio ns 2. w(i)t+1 = wt(i) t / (1­ t) 3. Re-estimate tree 4. Weight for each tree

· Final classification
­ Accuracy weighted vote across multip le trees

­ B = t / (1­ t)
· Where w(i)t = weight for the ith case in iteration t, and N is the total nu mber of cases
15


Post-Classification Processing
(Mc Iver and Friedl 2002, RSE)

· Application of Prior Probabilities
­ Glo ba l priors to remo ve tra in in g s ite class d istr ibution b iases ­ Mo vin g-w indow pr iors fro m earlier products ­ Use o f external maps of pr ior probab ilities to reso lve confus ions
· Agriculture/ natural vegetation co nfusion i n some regio ns · Use of city lights DMSP data to enha nce urban class accuracy

· Filling of Cloud-Covered Pixels from Earlier Maps
­ Use o f pre viou s year product when there are not suffic ient va lues to c lass ify a pixe l with confidence
16


Training Sites--STEP Database
(Mucho ney et al., 1999; PERS)

· STEP : ­ System for Terrestria l Ecosystem Parameterization ­ Interpreted from Landsat & ancillary data · Key STEP Parameters ­ Life form, co ver fraction, leaf type, pheno lo gy, ele vatio n, mo isture regime, d isturbance ­ Simp le description of s ite and type

A conf ide nce s ite near P ins k , Be lar us (20 x 2 0 k m)

17


IGBP Land Cover Units (17)
(Primary Layer) · Natural Vegetation (11)
­ Evergreen Needleleaf Forests ­ Evergreen Broadleaf Forests ­ Deciduous Needleleaf Forests ­ Deciduous Broadleaf Forests ­ Mixed Forests ­ Closed Shrublands ­ Open S hrublands ­ Woody Savannas ­ Savannas ­ Grasslands ­ Permane nt Wetlands

· De ve loped and Mosa ic Land s (3)
­ Croplands ­ Urban a nd B uilt-Up La nds ­ Cropland/Natural Vegeta tion Mosaics

· Non ve getated Lands (3)
­ Snow and Ice ­ Barren ­ Water Bodies

18


Global Sampling and STEP Maintenance
· Live (!!) Database: currently ~2300 sites globally


STEP Training Sites in Nelda Region
100
88

80

71 64

Number of training sites

63

60

45

40

37 30 28 28 22

20

18 11 7 11 12

0

1

3

4

5

6

7

8

9

10

11

12

13

14

16

17

IG BP class

20


IGBP site label and GLC2000
E GN 1
12 0 10 0

I GBP .1
1 20

D EN 3
10 0

I GBP .3
1 40

D EB 4
1 20

IG BP. 4

80

80

P i xel s co unt
1 3 4 5 6 7 9 12 18 26

P i xel s co unt

P i xel s co unt

60

60

40

40

20

20

1

4

5

6

12

0

0

0

20

40

60

80

10 0

3

GL C 20 00 cla ss

GL C 20 00 cla ss

GLC 2 00 0 class

21


IGBP site label and GLC2000
MIX 5
IG BP. 5
25
6 5

CSH 6
I GBP .6
50

OSH 7
I GBP .7

20

Steppe
P i xel s co unt 30

4

40

P i xel s co unt

P i xel s co unt

Barren tundra
20

15

10

2

3

0

0

1

1

3

4

5

6

7

3

11

18

20

21

0

10

5

15

17

18

23

24

GLC 2 00 0 class

GL C 20 00 cla ss

GL C 20 00 cla ss

22


IGBP site label and GLC2000
12

IGBP Sa vannas
Shrub tundra

IGBP .9

8

10

Pixels count

Steppe
6 0 2 4

1

3

7

11

12

13

18

21

23

GLC2000 class


IGBP sites label and GLC2000
I GB P.1 0
50

IGB P.1 1
250

IGB P.1 2

40

30

35

P ix el s co u nt

30

Pi xe l s co un t

25

Pi xe l s co un t
1 5 12 13 14 26

20

15

20

10

10

0

0

5

7

11

12

17

18

20

21

23

24

GL C 2 0 00 cla ss

0

50

100

15 0

2 00

1

3

7

10

11

12

20

21

22

23

GL C 20 0 0 cl as s GL C 20 0 0 cl as s

24


IGBP sites label and GLC2000
IG BP .1 3
100
8

I GBP .1 4
25

I GBP .1 6

80

60

6

P i xel s c ount

P i xel s count

40

P i xel s c ount

4

20

2

0

1

12

21

22

26

27

0

0

5

10

15

20

1

4

5

22

11

16

24

GLC 2000 c las s

G LC 2000 cl ass

G LC2000 cl ass

25


Proposed NELDA Land Cover Legend
Bas elin e Le gen d1 Tre e Do mina te d
Need leleaved Closed E vergr een Open3 Closed Deciduous Open Broadl eaved Closed E vergr een Open Closed Deciduous Open Cl osed Mi xed Op e n
1 2 3

Pos sible Add itiona l Distinctio ns

2

Cover Detail M ort al ity (yes/ no) S pecies W et land T rees (yes/no) Understory Char acteristics M anaged P lantation (T ree F ar m/Orchard)

T h e assu mp tio n is to u se h ig h r eso lu tio n ima ge ry (2 0 ­ 5 0 mete rs) an d min imu m map pi ng u n it 1 ­ 2 he cta res Clo se d >( > 6 5 ) % O p en (6 5 -1 5 )%

26


Proposed NELDA Land Cover Legend
Shrub D om inated Possible A dditional D istinctions

Closed Broadleaved Open Closed Needleleaved Open Closed Mixed Open

Species Wetland Shrubs ( yes/no) Leaf Longevity ­ Deciduous or Evergreen Tundra ( yes/no) Trees < 15 % Present/not Present (Trees < 5 %) Managed Plantations (Vineyard, for example) Tree Regeneration (yes/no)

27


Proposed NELDA Land Cover Legend
Baseline L egend P ossible Additional D istinctions Herbaceous Dominated

Closed Open

Speci es (grasses, l ichens, mosses, etc) Wetland Herb (yes/no) Tundra (yes/no) Pasture (yes/no)

Urban

Vegetat ion Dominated (Vegetati on Cover > 50 %) Non-Veget at ion Dominat ed (Vegetati on Cover < 50 %)

Bare Areas Permanent Snow and Ice Water
28


NELDA to LCCS
LC LP o s 1F o re s t 2 1F o re s t 2 1F o re s t 2 1F o re s t 2 2W o o d la n d 2 2W o o d la n d 2 2W o o d la n d 2 2W o o d la n d 2 3T hic ke t 2 3T hic ke t 2 4S hru b la n d2 4S hru b la n d2 5G ra s s la n d s 1 5G ra s s la n d s 1 1B uilt U p A1e a s r 1N a tura l W1 te rb a 0D ic h o to m 1 u s P o 2S no w 1 3Ic e 1 1F o re s t 2 1F o re s t 2 LC CCo d e LC 2 00 92 0 2 00 93 0 2 00 89 0 2 00 90 0 2 01 34 0 2 01 35 0 2 01 31 0 2 01 32 0 2 01 51 0 2 01 54 0 2 01 72 0 2 01 75 0 2 00 26 0 2 00 37 0 5 00 3-9 0 o8d i0 2 0e s 0 h a0s e 0 11 0 8 00 6 0 8 00 9 0 2 00 92 (2 )[Z 3] 0 2 00 92 (2 )[Z 4] 0 LCC Leev e lLCCO wnLLC C C Mo d abe A3A10 B2 XXD2 E 1 A3A10 B2 XXD2 E 2 A3A10 B2 XXD1 E 1 A3A10 B2 XXD1 E 2 A3A11 B2 XXD2 E 1 A3A11 B2 XXD2 E 2 A3A11 B2 XXD1 E 1 A3A11 B2 XXD1 E 2 A4A10 B3 XXD1 A4A10 B3 XXD2 A4A11 B3 XXD1 A4A11 B3 XXD2 A2A10 B4 A2A11 A4-A13 A1B 1 P re s B16 A2B 1 P re s A3B 1 P re s A3A10 B2 XXDNE 1.Z3MP re s T . 2 .E C. A3A10 B2 XXDNE 1.Z4BWre s T . 2 .E C. P O w nD e s c r l LCC La b e l Ma p C o d e N e e d le le a v e d E ve rg re e n T re e s 1 N e e d le le a v e d D e c id u o u s T re e s 2 B ro a d le a ve d E v e rg re e n T re e s 3 B ro a d le a ve d De c id uo us T re e s 4 N e e d le le a v e d E ve rg re e n W o o d la nd 5 N e e d le le a v e d D e c id u o u s W o o d la n6 d B ro a d le a ve d E v e rg re e n W o o d la n d 7 B ro a d le a ve d De c id uo us W o o d la n d8 B ro a d le a ve d S h rub s Clo s e 9 N e e d le le a v e d S hru b s C lo s e d 10 B ro a d le a ve d S h rub la nd 11 N e e d le le a v e d S hru b la n d 12 C lo s e d H e rb a c e o u s Ve g e ta tio n 13 H e rb a c e o u s O p e n Ve g e ta tio n 14 U rb a n Are a (s ) 15 P e re nn ia l N a tu ra l W a te rb o d ie s 19 B a re Are a (s ) 20 P e re nn ia l S no w 22 P e re nn ia l Ic e 23 e s (m o rta lity ) e e d le le a v e d E ve rg re e n T re e s N 24 tla nd N e e d le le a v e d E ve rg re e n T re e s 25

e n t > 11 m o nths e e e e n n n n t t c c > > e e 11 11 of of m o nths m o nths d e a d tre Bo g / W e

29


"Validation" Efforts
· Issues
­ ­ ­ ­ Lack of probab ility samp le M ixed p ixe l prob lem in coarse resolutio n data Amb iguo us clas s defin ition s Spectral separation of c lasses (can we actually distin gu ish them w ith MOD IS?)

· Approaches
­ Independent assess ments (Warren Cohen, OSU; Bigfoo t)
· NELDA sites for validation

­ Cross va lidation of STEP database Independent e va luation /assess ment activities ( independent e va luators) ­ Mode l-based assessment (confidences)
30


Cross Validation
(Strahler, 2003; http:// geo grap hy.bu.edu)

· Cross-Va lidation Procedure
­ Exploits STEP database ­ Hide 10 percent of training sites, classify with remaining 90 percent; repeat ten times for ten unique sets of all sites ­ Provides "confusion matrix" based on unseen pixels where whole training site is unseen ­ Not a stratified random sa mple, but a indication of accuracy
31


Summary
· MODIS Decis ion Tree · Add new examples from NELDA sit es to the STEP database · Review and change STEP polygons labels · Finalize NELDA legend

32