methods of validating maps of deforestation and selective logging carlos souza jr....
TRANSCRIPT
Methods of Validating Maps of Deforestation and Selective Logging
Carlos Souza [email protected]
Instituto do Homem e Meio Ambiente da Amazônia—ImazonCaixa Postal 5101, Belém, PA, Brasil CEP 66613-397
Department of Geography – University of CaliforniaSanta Barbara, California 93106, USA
[becky,matzke,carlos,mclark,numata,dar]@geog.ucsb.edu
Workshop Objectives
to propose protocols for ground-based accuracy assessments of remotely sensed estimates of deforestation and selective logged forests.
to address means of establishing error bars and accuracy estimates pertaining to deforestation and selective logging products.
Protocol for Accuracy Assessment of RS Products
Response DesignProtocol
Sampling DesignProtocol
Estimation andAnalysis Protocol
AccuracyAssessment
Stephen V. StehmanStephen V. StehmanRaymond L. CzaplewskiRaymond L. Czaplewski
RSE, 1998. 64:331-344RSE, 1998. 64:331-344
Sampling Protocol
Sample unit is a location (point) or
an area (pixel or polygon) used to assess the accuracy of a map.
links the map and the reference data. Pixel sampling
unit
Reference data (Ikonos)
Response DesignProtocol
Sampling DesignProtocol
Estimation andAnalysis Protocol
AccuracyAssessment
Stephen V. StehmanStephen V. StehmanRaymond L. CzaplewskiRaymond L. Czaplewski
RSE, 1998. 64:331-344RSE, 1998. 64:331-344
Sampling Protocol
Sampling method Requires an unbiased sample
Simple random Stratified random Cluster
Primary (psu) Secondary (ssu)
Polygon
psu
ssu
5x5 pixelblock
ssu
psu
““A scientifically defensible accuracy A scientifically defensible accuracy assessment requires a assessment requires a probability sampling”probability sampling”
Stephen V. StehmanStephen V. StehmanRaymond L. CzaplewskiRaymond L. Czaplewski
RSE, 1998. 64:331-344RSE, 1998. 64:331-344
Response DesignProtocol
Sampling DesignProtocol
Estimation andAnalysis Protocol
AccuracyAssessment
Response Design
It is a classification scheme applied to the sampling unit. Labeling protocol
Assigns a land-cover classification to the sampling unit based on the evaluation protocol:
Land cover proportion Predominant land-cover Centroid Multiple land cover classification Fuzzy classification
Evaluation:• 50% vegetation• 40% bare soil• 10% shrubs
Labeling:• Pasture
Evaluation:• pixel centerLabeling:
• Forest
Stephen V. StehmanStephen V. StehmanRaymond L. CzaplewskiRaymond L. Czaplewski
RSE, 1998. 64:331-344RSE, 1998. 64:331-344
Response DesignProtocol
Sampling DesignProtocol
Estimation andAnalysis Protocol
AccuracyAssessment
Reference Data
D – DecidiousC – ConiferAG – AgricultureSB - Shrub
43414111510375
10490374
1151985110
10085816
1152422465
D C AG SB Row Total
D
C
AG
SB
ColumnTotal
Overall Accuracy =(65+81+85+90)/434 = 74%
Cla
ssif
ied
Dat
a
D = 65/75 = 87%C = 81/103 = 79%AG = 85/115 = 74%SB = 90/141 = 64%
Producers Accuracy
D = 65/115 = 57%C = 81/100 = 81%AG = 85/115 = 74%SB = 90/104 = 87%
Users Accuracy
Source: Congalton and Green (1999)
Legend
Stephen V. StehmanStephen V. StehmanRaymond L. CzaplewskiRaymond L. Czaplewski
RSE, 1998. 64:331-344RSE, 1998. 64:331-344
Response DesignProtocol
Sampling DesignProtocol
Estimation andAnalysis Protocol
AccuracyAssessment
Estimation andAnalysis
Problems with the Standard Protocol
Study Area
Sampling Design
Flight-line randomly sampled by time-code Cluster samples centered on flight-line Reference sample unit = 30 m x 30 m area Five reference samples recorded per mosaic
Evaluation Protocol: Class Definitions Primary forest Pasture trees >20 m height herbaceous contiguous canopy shrubs < 2 m Second-growth forest annual crops shrubs > 2 m Urban/soil trees < 20 m human structures perennial crops roads Water bare soil
Response Design
Response Design Labeling protocol: dominant land-cover
34
1 2
0
Evaluation:100% Pasture
Label:“Pasture”
Evaluation:100% SecondaryForest
Label:“Secondary Forest”
Evaluation:55% Pasture45% Secondary Forest
Label:“Pasture”
Evaluation:100% Primary Forest
Label:“Primary Forest”
Accuracy AssessmentMethods and Analysis
Correct geolocation errors between video mosaics and map
Randomly select image frames from video based on time code
Mosaic and georectify video and overlay cluster samples
Group (i.e. five interpreters) re-interprets video to generate ‘gold’ reference data set
Five individuals independently interpret video to produce five reference data sets
Remove mixed samples and edge pixels; identify change
Parameters estimated from error matrices: overall accuracy, producer’s accuracy (omission errors), user’s accuracy (commission errors), Kappa (KHAT) statistics, Kappa variance (for statistical comparisons)
Interpreter vs. interpreter accuracyInterpreter vs. map accuracy
‘Gold’ reference vs. map accuracy
Influence of reference data corrections on map accuracy*statistically different Kappa coefficients (95% confidence level)
Version Corrections # Samples Total Agreement Kappa coefficient1 none 790 75.4% 0.60232 geocorrection 790 83.2% 0.7269
3geocorrection; remove change pixels 771 85.3% 0.7595
4geocorrection; remove mixed reference samples 684 85.4% 0.7437
5geocorrection; remove map edge pixels 605 88.1% 0.7914
6geocorrection; remove mixed reference and map edge pixels 532 91.0% 0.8283
7geocorrection; remove mixed, edge, and change pixels 523 92.5% 0.8557
Area estimation accuracy assessment
Gross Deforestation in the State of Acre 1999 a 2002
0.0
2000.0
4000.0
6000.0
8000.0
10000.0
12000.0
14000.0
16000.0
18000.0
1999 2000 20001 20002
Area (
km
2 )
Imazon-Imac
INPE
33%21%
23%17%
Average = 23.50%
Stdev = 6.81%
Deforestation Estimates in Acre
Area of Forest Fragments:
2864.9 km2
Are the Fragments “Capoeiras”?Deforestation Age
Forest
Deforestation until 2000
Water
Forest – Prodes digital
Landsat Image Classification (Roberts et al., 2003)
FragmentsFragments
Classification (INPE - Prodes Digital)
Scale Effect on Mapping Deforestation
1:50.000
1:250.000
9 km
9 km
1:50.000
1:250.000
Scale Effect on Mapping Deforestation
Deforestation by Classes of Size
Desmatamento - 2003
9720
14139
7348
4272
29582022
44083481
1727598 858
0%
20%
40%
60%
80%
100%
0.36-5 5-10 10-15 15-20 20-25 25-30 30-50 50-100 100-200
200-300
>300Área (ha)
Pe
rce
ntu
al
de
po
líg
on
o
de
sm
ata
do
0
3000
6000
9000
12000
15000
N°
de
po
lig
on
o
de
sm
ata
do
Desmatamento 2002
8794
11550
6306
3710
25611959
42503753
2030
737 9860%
20%
40%
60%
80%
100%
0.36-5 5-10 10-15 15-20 20-25 25-30 30-50 50-100 100-200
200-300
>300
Área (ha)
Pe
rce
ntu
al
de
po
lig
on
o
de
sm
ata
do
0
3000
6000
9000
12000
15000
N°
de
po
lig
on
o d
es
ma
tad
o
Contribição do polígono desmatado brutoContribuição cumulativa do polígono desmatadoNúmero de polígonos desmatado
Co
ntr
ibu
içã
o p
/ o
D
es
ma
tam
en
to (
%)
Co
ntr
ibu
içã
o p
/ o
D
es
ma
tam
en
to (
%)
Nú
me
ro d
e P
olí
go
no
sN
úm
ero
de
Po
líg
on
os
Classification Errors in 2002
Old deforestation
Forest
Increment
Incremento 2002 - INPE
Classification Imazon/Imac Classification - Prodes
Accuracy estimates of deforestation and selective logging products
Reference DataHigh resolution satellite image
Ikonos
Field data Several sources
Ikonos Data for LBA-Ecology Study Sites
Hurtt et al., (2003)
Proposal Map accuracy
Sample design: random cluster samplingResponse design: ?Estimation design: ?Correct reference data errorsSeveral research groups interpreting the
reference data Include area estimation accuracy