ESA-GLOBCOVER - A Global Land Cover Service at 300 m from ENVISAT MERIS Time Series

Transcription

1 October 15, 2008 (Jena Germany) ESA-GLOBCOVER - A Global Land Cover Service at 300 m from ENVISAT MERIS Time Series Defourny P.1, Bontemps S., Vancutsem C.1, Pekel J.F.1,, Vanbogaert E.1, Bicheron P.2, Leroy M.2, Brockmann C.3, Krämer U.3, Schouten L. 4, Arino O. 5 UCL-Geomatics (Belgium) 2 Medias-France Brockmann Consult (Germany) 4 INFRAM (The Netherlands) 5 ESA 1 3 October 15, 2008 (Jena Germany) 1

2 an ESA initiative (Coord.: O.Arino) carried out by October 15, 2008 (Jena Germany) 2

3 GLOBCOVER Objectives Develop and demonstrate a service of production of a global land cover map for 2005/2006 at 300 m resolution using MERIS FR based on FAO Land Cover Classification System Two deliverables A hardware & software system Two validated products Time composited surface reflectances Land cover map October 15, 2008 (Jena Germany) 3

4 GLOBCOVER Challenges 3 ENVISAT-MERIS FR Challenges Continuous acquisition of MERIS FR data set at global scale Geolocation of MERIS FR data set within a pixel Very large volume of data to process 4 Land Cover Classification Challenges Automatic and repeatable processing chain Consistent at global scale from MERIS data Building on GLC2000 experience Legend documented using FAO Land Cover Class. System October 15, 2008 (Jena Germany) 4

5 MERIS FR data acquisition Acquisition constraints No systematic acquisition in FR mode Max. acquisition to be raised Artemis mask Outside Artemis mask, use of recorder (+ recently local receiving stations) => 19 months acquisition period (Dec. 05 Jun.06) => MERIS L1 product (> 20 TByte) 15 bands at 300 meters resolution (VIS-NIR from 390 nm to 1040 nm) swath width of 1150 km initial geolocation accuracy specification ~ 2 km! October 15, 2008 (Jena Germany) 5

6 MERIS FR Data Acquisition October 15, 2008 (Jena Germany) 6

7 MERIS FR Data Acquisition Number of valid observations for the 19 months October 15, 2008 (Jena Germany) 7

8 Improved MERIS FR Pre-Processing Dedicated Atmospheric Correction using MERIS RR aerosols Cloud and Shadow Screening AMORGOS (ACRI) + Projection Tools for a much improved Geolocation Accuracy: relative RMSE = 51.6 m (ca pixel) absolute RMSE = 77.1 m (ca pixel) Compositing Strategy (Vanctusem et al., CJRS 2008) October 15, 2008 (Jena Germany) 8

9 MERIS FR composite Spain/Morocco Seasonal composite (Sep-Nov) over a 10 x 10 area October 15, 2008 (Jena Germany) 9

10 MERIS FR composite Nile Valley Monthly composite (May) over a 2.5 x 2.5 area October 15, 2008 (Jena Germany) 10

11 GLOBCOVER Classification concept Key idea: Combine the high spatial consistency of classes delineation obtained from multispectral composite(s) with the great land cover discrimination from temporal profiles analysis GLOBCOVER Principles: Regionally-tuned approach based on 22 equal-reasoning areas Multispectral composites and reflectance time series Typology defined and documented using FAO LCCS and as much as possible compatible with GLC2000, EEA CORINE Land Cover product and, Africover.. International experts inputs for the classification algorithm calibration October 15, 2008 (Jena Germany) 11

12 Classification strategy Each region is processed independently using regionally tuned classif. parameters Step 0: A priori stratification Split the world in 22 equal-reasoning regions from ecological and remote sensing point of view Step 1: For each region, per-pixel pixel sup./unsup unsup.. classification algorithm Spatially consistent spectral cluster Step 2: Per-cluster temporal characterization Robust temporal metrics computed at cluster level from biweekly multispectral time series and associated indices Step 3: Per-cluster classification algorithm Consistent unlabelled spectro-temporal classes Step 4: Labelling rule-based procedure Based on best existing products and experience of an international expert network to get the LCCS land cover classes Step 5: Calibration MERIS specific labelling rules thanks to interactive calibration by a network of international experts Step 6: Independent Validation Validation Land Cover product accuracy October 15, 2008 (Jena Germany) 12 n classes R (%) 4,5 4 3,5 3 2,5 2 1,5 1 0,5 x classes NDVINDVI Experts clustering temps Vegetation Expert & Ancillary data

13 GLOBCOVER Stratification 22 equal-reasoning regions from ecological and remote sensing point of view October 15, 2008 (Jena Germany) 13

14 GLOBCOVER stratification for regionally tuned discrimination June - August ST09 ST07 Sept.-November April-June ST11 July-August April - May October 15, 2008 (Jena Germany) 14

15 Per-pixel classification providing consistent spectral clusters LUT1 + training dataset 1. Supervised classification on poorly represented classes MASK Annual, seasonal or bimonthly MERIS composites MASK LUT2 2. Unsupervised classification on unclassified areas N classes October 15, 2008 (Jena Germany) 15

16 Per-cluster classification from spectrotemporal characterization Robust temporal metrics computed per-cluster from biweekly multispectral composite and associated indices MERIS Reflec tance or index Cluster attributes : Min., Max., Multispectral reflectance Statistical profile analysis Spectro-temporal classes October 15, 2008 (Jena Germany) 16

17 Automated labelling procedure based on best existing LC maps Reference dataset based on existing LC data: The GLC2000 global and regional maps over the world ( The Corine LC map 2000 (EA) The Africover map over 10 countries of Africa (FAO, The National Land Cover Database (NLCD) of United States (USGS) The National Land Cover map 2000 of China (Chinese Academy of Sciences) The Australia land use map (Australian Government) The Democratic Republic of Congo map (UCL-Geomatics, 2006, The land use classification of the agricultural areas of Argentina (Facultad de Agronomía, Universidad de Buenos Aires) The land cover database of Burkina-Faso (BDOT project, IGN France International) The land cover classification of Mexico (Giri and Jenkins, 2005) The land cover map of Canada 2005 (Canada Centre for Remote Sensing) The land cover map JICA of Cambodia (2000) October 15, 2008 (Jena Germany) 17

18 Rainfed Cropland GLOCOVER Legend 22 LCCS classes Post-flooding or irrigated croplands Mosaic cropland (50-70%) / vegetation (grassland/shrubland/forest) (20-50%) Mosaic vegetation (grassland/shrubland/forest) (50-70%) / cropland (20-50%) Closed to open (>15%) broadleaved evergreen and/or semi-deciduous forest (>5m) Closed (>40%) broadleaved deciduous forest (>5m) Open (15-40%) broadleaved deciduous forest/woodland (>5m) Closed (>40%) needle-leaved evergreen forest (>5m) Open (15-40%) needle-leaved deciduous or evergreen forest (>5m) + sub-classes at regional level Mosaic grassland (50-70%) and forest or shrubland (20-50%) Closed to open (>15%) shrubland (<5m) Closed to open (>15%) grassland Sparse (<15%) vegetation Closed (>40%) broadleaved forest regularly flooded, fresh water * Closed (>40%) broadleaved semi-deciduous and/or evergreen forest regularly flooded, saline water Closed to open (>15%) grassland or shrubland or woody vegetation on regularly flooded or waterlogged soil, fresh, brackish Artificial or saline surfaces water and associated areas (Urban areas >50%) Bare Areas Closed to open (>15%) mixed broadleaved Water Bodies and needleaved forest Permanent Snow and Ice Mosaic forest or shrubland (50-70%) and October 15, 2008 (Jena Germany) 18 grassland (20-50%) No Data * Resulting from the reference dataset

19 Calibration Phase Step 0: A priori stratification Split the world in 22 equal-reasoning regions from ecological and remote sensing point of view Step 1: For each region, per-pixel pixel sup./unsup unsup.. classification algorithm Spatially consistent spectral cluster Step 2: Per-cluster temporal characterization Robust temporal metrics computed at cluster level from biweekly multispectral time series and associated indices Step 3: Per-cluster classification algorithm Consistent unlabelled spectro- temporal classes Step 4: Labelling rule-based procedure Based on best existing products and experience of an international expert networkusing to get the LCCS land cover classes Step 5: Calibration MERIS specific labelling rules thanks to interactive calibration by a network of international experts Step 6: Independent Validation Land Cover product accuracy October 15, 2008 (Jena Germany) 19 n classes R (%) 4,5 4 3,5 3 2,5 2 1,5 1 0,5 x classes NDVINDVI Experts clustering temps Vegetation Expert & Ancillary data

20 GLOBCOVER Land cover V2 based on 19 months of acquisition ( ) October 15, 2008 (Jena Germany) 20

21 GLOBCOVER Land cover V2 much better spatial resolution Reference Globcover V2 October 15, 2008 (Jena Germany) 21

22 GLOBCOVER Land cover V2 better thematic discrimination Reference Globcover V2 October 15, 2008 (Jena Germany) 22

23 GLOBCOVER Land cover V2 builds on already existing LC map Corine LC Globcover V2 October 15, 2008 (Jena Germany) 23

24 Corine LC Compatibility with Corine LC? thematic error > change rate Globcover V2 October 15, 2008 (Jena Germany) 24

25 Global land cover area estimate October 15, 2008 (Jena Germany) 25

26 Annual / seasonal composites variable quality depending on data availability October 15, 2008 (Jena Germany) 26

27 Annual / seasonal composites variable quality depending on data availability October 15, 2008 (Jena Germany) 27

28 GLOBCOVER Validation process according to CEOS Guidelines (2006) Independent Repeatable Transparent Scientifically sound Acceptable for the LC community > Stratified sampling based on equal area projection October CDR Review 15, 2008 Toulouse (Jena March Germany) 15th

29 Validation points described according to LCCS hierarchical structure October 15, 2008 (Jena Germany) 29

30 GLOBCOVER Validation tool Google / Virtual Earth LCCS Classifiers+ Level of Certainty 8-y SPOT VGT NDVI profiles October 15, 2008 (Jena Germany) 30

31 Reference data source : 6 worshops involving 16 international LC experts October 15, 2008 (Jena Germany) 31

32 4258 validation points based on HiRes & temporal profiles interpretation by experts Blue points: Globcover project (3917 points) (including 225 double interpretation by 2 experts) Brown points: IMWI data October 15, 2008 (Jena Germany) 32

33 GLOBCOVER validation data set From the 3167 points considered as certain, heterogeneous points were removed (cfr GLC2000) dominance not considered as an error forest point with no phenology specified assigned as Mixed forests 2115 homogeneous validation points to validate the GLOBCOVER v.2 at global legend level (22 classes) GOFC-GOLD Second End Users Land Cover Meeting Symposium Rome, October March 10, 15, 2008 (Jena Germany) 33

34 GLOBCOVER overall accuracy : 73,1 % GLOBCOVER LC product (22 classes) versus certain and homogeneous validation points : 67,6 % (n = 2215) GLOBCOVER LC product (22 classes) versus certain validation points & LC class area weighting : 73,1 % (n = 3167) October 15, 2008 (Jena Germany) 34

35 Consistency between experts at dominant land cover level (75,7 %) 225 points with interpretation by 2 independent experts October 15, 2008 (Jena Germany) 35

36 ESA Globcover Data access tool Available 10 bi-monthly Reflectance 1 annual Reflectance Land Cover + PDM (Sept. 30) + Validation Report (Oct. 30)! October 15, 2008 (Jena Germany) 36 Postel/Medias-France GlobCover Data access tool ESA Zoomify tool for GLOBCOVER product consultation

37 GLOBCOVER v.2 map more than a 300 m MERIS map October 15, 2008 (Jena Germany) 37

38 October 15, 2008 (Jena Germany) 38

39 Thank for your attention October 15, 2008 (Jena Germany) 39