GLC 2000 ‘Final Results’ Workshop(JRC-Ispra, 24 ~ 26 March, 2003)
GLC 2000 ‘Final Results’ Workshop(JRC-Ispra, 24 ~ 26 March, 2003)
LAND COVER MAP OF FRANCE USING S1/S10 SPOT/VEGETATION DATA
LAND COVER MAP OF FRANCE USING S1/S10 SPOT/VEGETATION DATA
Jean-Louis CHAMPEAUXKyung-Soo HAN
SPOT4-VEGETATION1 DATASPOT4-VEGETATION1 DATA
S1 Daily Synthesis Products– 4 Spectral Channels (b0, b2, b3, SWIR)
– Angular Information (SZA, VZA, SAA, VAA)
Period– 1 January ~ 31 December, 2000
– 351 Daily Synthesis Sets (15 missing days)
Zone
5.50 W ~ 9.91 E
51.50 N ~ 40.01 N
METHODOLOGY SYNTHESISMETHODOLOGY SYNTHESIS
Spot4-VGT1S1 Daily Synthesis
Cloud Mask
SWIR BlindSuppression
BRDFCorrection
ReflectanceNormalization
Principal ComponentsTransformation
K-mean ClusteringAlgorithm
Map of Unnamed Classes
CORINE Land Cover
ConfusionMatrix
New LandCover MapNew LandCover Map
AngularInformation
Image Treatment CLOUD MASKCLOUD MASK
New Tuned Thresholdsfor Surface Reflectance
b0 220
SWIR 180
CLOUDY CLEARSKY
Yes
Yes
No
No
b0, 23/08/2000
b0, 5/02/2000
SWIR BLIND SUPPRESSION-case1SWIR BLIND SUPPRESSION-case1
Image Treatment
Core(defect)
dilatation
Core
–by threshold
Dilatation– Ps(i-1,j) and Pe(i+1,j) Ps(i,j): start pixel of core on a line Pe(i,j): end pixel of core on a line
SWIR BLIND SUPPRESSION-case2SWIR BLIND SUPPRESSION-case2
Image Treatment
(SWIR-b2)2
= (SWIR-b2)2
=
Threshold = M’ + (M 3.0)M’ : Mean of M values for a lineM : Standard dev of M values for a line
SEMI-EMPIRICAL BRDF MODELSEMI-EMPIRICAL BRDF MODEL
Image Treatment
s, v, = k0 + k1 f1s, v, + k2 f2s, v, , Roujean et al., 1992
N measured
N: number of clear days during a compositing period
k0: bidirectional reflectance, s = v = 0
Inversio
n
of
system
k0, k1, k2
Computati
on of
normalized
reflectanc
e
NORMALIZATION OF NORMALIZATION OF
Image Treatment
Method (Duchemin & Maisongrande, 2002):norm,i= model(s=moyen, v=0) + mesured,i model,i(s, v,)
S1-Syn.(RGB)19 June 2002R: SWIRG: b3B: b2
Norm. (RGB)19 June 2002R: SWIRG: b3B: b2
Day of Year1 10 20 30 40
Composite 1
Composite 2
Composite 3::
Average all norm values for clear day
- 31-day Screening- > 4 Observations
n
iinormN
n 1)(
1
N
NORMALIZATION OF NORMALIZATION OF
Image Treatment
norm(i): norm of day i
N: composite value by N norm
N: number of clear days for the compositing period
EXAMPLES OF 10-day COMPOSITEEXAMPLES OF 10-day COMPOSITE
Image Treatment
1st 10-day 8th 10-day 12th 10-day
17th 10-day 34th 10-day23 10-day
23th 10-day ( August)
TIME SERIES PROFILE TIME SERIES PROFILE
Image Treatment
AREA OPTIMIZATIONAREA OPTIMIZATION
Memory limitations to run the algorithm over the whole areaMemory limitations to run the algorithm over the whole area
Reduction of running size
Step1.Select a zone through a climate map
Step2.Divide into two parts to avoid
mosaic problems
North part
South part
Classification
PRINCIPAL COMPONENTSPRINCIPAL COMPONENTS
Classification
North area465,942 pixels26 10-days
(78 input images)
36 10-day images for each channel (=117 variables)
Select images by % of cloudy pixels over each area
South area517,314 pixels
24 10-days(72 input images)
PRINCIPAL COMPONENTS ANALYSIS
Selection of ComponentsComponents having 99% accumulated correlation
North area48 components
South area34 components
K-MEAN CLUSTERINGK-MEAN CLUSTERING
Classification
North area48 components
South area34 components
K-mean Clustering for 40 classes
1st output (40 classes)
2nd output (44 classes)- 40 classes from K-mean clustering- 4 classes from CORINE
CORINE Mask for misdetected pixels
- Artificial surfaces- Water bodies- Wetlands- Beachs, dunes, sandsCORINE MASK
Agglomeration of classes fitting
the same landcover
MOSAICMOSAIC
Classification
North partMap
North partMap
South partMap
South partMap
Mosaic
MISSING
MISSING PIXEL TREATMENTMISSING PIXEL TREATMENT
Classification
Missing due tocloud
Missing due to cloud & snow
Missing due to cloud & snow
K-mean clusteringfor missing pixels
A
B C
With a clear imageafter the end of snow melting
- Julian day 254
NEW URBAN DETERMINATIONNEW URBAN DETERMINATION
Classification
A Supervised Classification for New Urban Area Class
Toulouse Toulouse New Urban
Artificial Surfaces Mask from CORINE 1992
- Shrub Land- Sparsely Vegetated Area
Artificial Surfaces Mask from CORINE 1992
Initial classification
After re- classification
FINAL RESULTFINAL RESULT
Classification
VALIDATIONVALIDATION
Reference: CORINE
EUROPEglc2000
MODIS
PELCOM
FRANCEglc2000
Confusion Matrix
Com
pari
son
CONFUSION MATRIX WITH CORINECONFUSION MATRIX WITH CORINE
Validation
Accuracy =Correctly Detected Pixels
Total Detected Pixels from reference
FRANCE* : Grasslands & Forest+Pastures
66.8112.3056.5616.07Barren Land
43.4337.115.5642.7039.75Grasslands
39.03-27.7135.42Permanent Crop
80.7991.1880.7480.43Arable Land + Perm. Crop
80.2680.4880.5980.10Arable Land
17.4919.721.6914.74Shrubland
77.8247.1773.2370.77Forest
FRANCE
MODISEUROP
EPELCO
MFRANCE
*% (Accuracy 100)
CONFUSION MATRIX WITH CORINECONFUSION MATRIX WITH CORINE
Validation
Reliability =(User’s accuracy)
Correctly Detected Pixels
Total Detected Pixels in the classified data
FRANCE* : Grasslands & Forest+Pastures
72.3192.5180.8988.40Barren Land
50.4357.7240.2352.1748.68Grasslands
52.67-55.7458.94Permanent Crop
60.4241.0051.7447.26Arable Land + Perm. Crop
63.1448.9153.1548.60Arable Land
35.417.0820.6437.01Shrubland
67.1349.9560.7352.26Forest
FRANCE
MODISEUROP
EPELCO
MFRANCE
*% (Reliability 100)
CONFUSION MATRIX WITH CORINECONFUSION MATRIX WITH CORINE
Validation
Comparison Index (CI) = (Reliability Accuracy)0.5
FRANCE* : Grasslands & Forest+Pastures
0.700.340.680.38Barren Land
0.460.460.150.470.44Grasslands
0.450.090.390.46Permanent Crop
0.690.610.640.61Arable Land + Perm. Crop
0.710.630.650.62Arable Land
0.250.120.060.23Shrubland
0.720.490.670.61Forest
FRANCE
MODISEUROP
EPELCO
MFRANCE
*0 -1
Reliability
CI
CONFUSION MATRIX WITH CORINECONFUSION MATRIX WITH CORINE
Validation
11.5413.7915.4712.45Mixed Forest
46.0557.5942.1130.60Coniferous Forest
49.3417.6733.3834.50Broad-leved Forest
FRANCE
MODISEUROP
EPELCO
M
12.4034.6415.0621.86Mixed Forest
63.3925.6851.4541.52Coniferous Forest
51.3313.5242.4342.08Broad-leved Forest
Accuracy
0.120.220.150.17Mixed Forest
0.540.390.470.36Coniferous Forest
0.500.160.380.38Broad-leved Forest
Forest…in
detail
Overall Accuracy
Total
%
0 10 20 30 40 50 60 70 80 90 100
%
0 10 20 30 40 50 60 70 80 90 100
FRANCEMODISEUROPEPELCOM
CI x 100
Accuracy
Reliability50.94
57.51
59.83
61.7664.96
70.00
0.560.61
0.650.45
50.44
45.00
CONFUSION MATRIX WITH CORINECONFUSION MATRIX WITH CORINE
Validation
Improvement of the classification due to the use of reflectances (B2,B3,SWIR) instead of NDVI
CONCLUSIONSCONCLUSIONS
Production of consistent normalized 10-day composites
Determination of new urban increase
Improvement of regional classification compared to the whole Europe (glc2000), PELCOM and MODIS products