30 Meter Vegetation Continuous Fields for the Continental...
1
30 Meter Vegetation Continuous Fields for the Continental United States Conclusions The opening of the Landsat archive affords researchers the possibility for producing large area land cover extent and change products. The experience gained using AVHRR and MODIS data to depict large area land cover is now being directly applied to Landsat imagery. The results shown here indicate the potential for operationally implementing VCF characterization methods at national scales. A global implementation is feasible. However, the United States is the most data-rich environment for Landsat, as all overpasses are acquired. The extrapolation of this method to other regions will have to adjust to less frequent acquisitions. The derivation of a standard method for global VCF mapping is forthcoming. Primary Reference • HANSEN, M.C., EGOROV, A, ROY, D.P., POTAPOV, P., JU, J., TURUBANOVA, S., KOMMAREDDY, I., LOVELAND, T. , 2011, Continuous fields of land cover for the conterminous United States using Landsat data: First results from the Web-Enabled Landsat Data (WELD) project. Remote Sensing Letters, 2, 4:279-288. Other References • BREIMAN, L., 1996, Bagging predictors. Machine Learning, 26, pp. 123-140. • BREIMAN, L., FRIEDMAN, J., OLSHEN, R., STONE, C., 1984, Classification and Regression Trees (Wadsworth and Brooks/Cole, Monterey, CA). • FRIEDL M.A., MCIVER, D.K., HODGES, J.C.F., ZHANG, X.Y., MUCHONEY, D., STRAHLER, A.H., WOODCOCK, C.E., GOPAL, S., SCHNEIDER, A., COOPER, A., BACCINI. A., GAO, F., and SCHAAF, C., 2002, Global land cover mapping from MODIS: algorithms and early results. Remote Sensing of Environment , 83, pp. 287-302. • HANSEN, M., DUBAYAH, R., and DEFRIES, R., 1996, Classification trees: An alternative to traditional land cover classifiers. International Journal of Remote Sensing, 17, pp. 1075-1081. • HANSEN, M. C., DEFRIES, R. S., TOWNSHEND, J. R. G., CARROLL, M., DIMICELI, C., and SOHLBERG, R. A., 2003, Global percent tree cover at a spatial resolution of 500 meters: First results of the MODIS vegetation continuous fields algorithm. Earth Interactions, 7, paper no. 10, 15 pp. [online journal] • ROBERTS, D. A., Gardner, M., Church, R., Ustin, S., Scheer, G. and Green, R. O., 1998. Mapping Chaparral in the Santa Monica Mountains using multiple endmember spectral mixture models. Remote Sensing of Environment , 65, pp. 267−279. Change analysis Tree cover Coastal plain, Bertie county, NC, N 36.1344° W 77.1074° no change increase in tree cover decrease in tree cover Bare ground Dallas – Fort Worth, TX, N 32.8694° W 97.0481° no change increase in bare ground decrease in bare ground Introduction Systematic large area land cover monitoring at medium spatial resolution has to date been limited primarily by data costs and processing limitations. However, since January 2008, radiometrically and terrain corrected Landsat data have been made available free of charge over the internet by the USGS Center for Earth Resources Observation and Science (EROS) (Woodcock et al. 2008), and the cost of computer processors and data storage continues to decline. A NASA-funded project, entitled Web-Enabled Landsat Data (WELD), has demonstrated recently the potential of large volume Landsat 7 ETM+ data processing to provide 30m spatially mosaiced and temporally composited timeseries data sets for the conterminous United States (CONUS) and Alaska (http://landsat.usgs.gov/WELD.php) (Roy et al. 2010). This project presents the application of the MODIS (Moderate Resolution Imaging Spectroradiometer) Vegetation Continuous Fields (VCF) characterization approach (Hansen et al. 2003) to the WELD data to produce 30m annual CONUS land cover estimates of percent tree, percent other vegetation (including grasses, shrubs and forbs) and percent bare ground cover, and also the probability of water presence and percent of snow and ice. Methods Tree-based classification methods are well-established in the remote sensing community and have been used to generate the MODIS Vegetation Continuous Field and Land Cover products (Hansen et al. 2003; Friedl et al. 2002). In this research, a bagged decision tree methodology was used to classify the WELD data. Decision trees are hierarchical classifiers that predict class membership by recursively partitioning a data set into more homogeneous subsets, referred to as nodes (Breiman et al. 1984; Breiman 1996). Trees can accept either categorical data in performing classifications (classification trees) or continuous data in performing sub-pixel percent cover estimations (regression trees). For regression trees, a sum of squares criterion is used to split the data into successively less varying subsets. The split that yields the maximum reduction in variability is selected to build the regression tree model. For annual estimates of percent tree cover and bare ground, regression tree models were employed using a bagging methodology. For preliminary weekly percent water and snow/ice, we applied a linear mixture model, as sub-pixel percent training data were not available. The method is that of Roberts et al. (1998), which employs multiple end-members to linearly model percent cover per pixel. These estimates can then be used as training data for regression tree model implementation. Deriving percent cover training data using high spatial resolution imagery Visual crown/no crown classifications of thirty-two 4 m Ikonos and three 2.8 m QuickBird multi-spectral images distributed across the CONUS were used to derive sub-Landsat pixel percent crown cover data based on the method described by Hansen et al. (2002) Percent crown cover per 30m WELD grid cell was then calculated and used as training data. Secondly, an additional 3,295,525 30m pixels of either 100% or 0% canopy cover were interpreted visually using very high spatial resolution (<10 m) Google Earth imagery. These additional training pixels were distributed in 2,156 homogeneous polygons across the CONUS and were checked by comparison with the 2008 WELD data to ensure that there was no change between 2008 and the acquisition of the GoogleEarth imagery. Visual interpretation of Google Earth imagery was performed on an ad hoc basis to fill regional gaps where very high spatial resolution Ikonos and QuickBird imagery were lacking. X – a pixel of mixed class 100 * ' 2 ' 1 ' 1 % d d d X Flowchart of classification process Metrics Sampling (25 samples) Building models (25 regression trees) Georanking Median (rank 13) as classification result Training data WELD hdf web storage ftp://weldftp.cr.usgs.gov Auxiliary datasets Classification (25 preliminary classification results) Deriving training using linear spectral unmixing Spectral Mixture Analysis was used to create training data for snow/ice and water layers. Nearest neighbors from pure classes (0% and 100%) were used as end-members to calculate percent snow/ice and water for pixels of mixed class (1..99%) (Roberts et al., 1998). Euclidean distance measure of closeness was used for this calculation. Band 5 Band 4 Band 3 Multiyear growing season composite image trees other vegetation bare ground water 2008 VCF layers as r-g-b composite Snow and Ice cover Croplands near to Hermiston, OR, N 45.9012°, W 119.368° Winter mosaic. Bands 5,4,3 as RGB 0% % of snow and Ice cover 100% 0% % of open water 100% Open water surface Marshes close to Sea Breeze, NJ, N 39.3071° W 75.2616° Summer mosaic. Bands 5,4,3 as RGB Weekly updated layers X Y Z Y – nearest train pixel of 100% pure class d2 – distance from X to Y Direct line connecting 2 and-members (Y and Z) d1’ d2’ d1’ and d2’ are the projections of d1 and d2 Spectral space Z – nearest train pixel of 0 % pure class 90% percentile 75% percentile 50% percentile 25% percentile 10% percentile Band2 TOA Reflectance Band3 TOA Reflectance Band7 TOA Reflectance Band62 TOA Brightness temperature Band61 TOA Brightness temperature Band5 TOA Reflectance Band4 TOA Reflectance Band1 TOA Reflectance d1 Minimum bare ground within the growing season Tree canopy density within the growing season New York, NY, NJ, N 40.7628°, W 73.9825° Multiyear composite. Bands 5,4,3 as RGB. 0% % of bare ground 100% Mount Rainier, WA, N 46.8519°, W 121.774° Multiyear composite. Bands 5,4,3 as RGB. 0% % of tree cover 100% 2006 2007 2008 2009 2010 2006 2007 2008 2009 2010 Annually updated layers Tree cover > 10,000,000 training pixels derived mostly from crown/no crown classifications of IKONOS and QuickBird images Bare ground > 150,000,000 training pixels from NLCD impervious surface and MODIS VCF layers Water > 130,000 visually interpreted training pixels with following spectral unmixing Snow/ice > 50,000 visually interpreted training pixels with following spectral unmixing Tree cover IKONOS and QuickBird images, Google Earth Bare ground MODIS VCF, SRTM water bodies and NLCD impervious surface layers Water Google Earth Snow/ice Google Earth Alexey V. Egorov, Matthew C. Hansen, Peter Potapov, Svetlana Turubanova, David P. Roy, Junchang Ju - South Dakota State University, Brookings, SD NASA MEASURE program grant number NNX08AL93A d1 – distance from X to Z d2 100% 1…99% 0% 0% 100% 0% 100% 0% 100%
Transcript of 30 Meter Vegetation Continuous Fields for the Continental...
30 Meter Vegetation Continuous Fields for the Continental United States
ConclusionsThe opening of the Landsat archive affords researchers the possibility for producing large area land cover extent and change products. The experience gained using AVHRR and MODIS data to depict large area land cover is now being directly applied to Landsat imagery. The results shown here indicate the potential for operationally implementing VCF characterization methods at national scales. A global implementation is feasible. However, the United States is the most data-rich environment for Landsat, as all overpasses are acquired. The extrapolation of this method to other regions will have to adjust to less frequent acquisitions. The derivation of a standard method for global VCF mapping is forthcoming.
Primary Reference• HANSEN, M.C., EGOROV, A, ROY, D.P., POTAPOV, P., JU, J., TURUBANOVA, S., KOMMAREDDY, I., LOVELAND, T. , 2011, Continuous fields of land cover for the conterminous United States using Landsat data: First results from the Web-Enabled Landsat Data (WELD) project. Remote Sensing Letters, 2, 4:279-288.
Other References• BREIMAN, L., 1996, Bagging predictors. Machine Learning, 26, pp. 123-140.• BREIMAN, L., FRIEDMAN, J., OLSHEN, R., STONE, C., 1984, Classification and Regression Trees (Wadsworth and Brooks/Cole, Monterey, CA).• FRIEDL M.A., MCIVER, D.K., HODGES, J.C.F., ZHANG, X.Y., MUCHONEY, D., STRAHLER, A.H., WOODCOCK, C.E., GOPAL, S., SCHNEIDER, A., COOPER, A., BACCINI. A., GAO, F., and SCHAAF, C., 2002, Global land cover mapping from MODIS: algorithms and early results. Remote Sensing of Environment, 83, pp. 287-302.• HANSEN, M., DUBAYAH, R., and DEFRIES, R., 1996, Classification trees: An alternative to traditional land cover classifiers. International Journal of Remote Sensing, 17, pp. 1075-1081.• HANSEN, M. C., DEFRIES, R. S., TOWNSHEND, J. R. G., CARROLL, M., DIMICELI, C., and SOHLBERG, R. A., 2003, Global percent tree cover at a spatial resolution of 500 meters: First results of the MODIS vegetation continuous fields algorithm. Earth Interactions, 7, paper no. 10, 15 pp. [online journal]• ROBERTS, D. A., Gardner, M., Church, R., Ustin, S., Scheer, G. and Green, R. O., 1998. Mapping Chaparral in the Santa Monica Mountains using multiple endmember spectral mixture models. Remote Sensing of Environment, 65, pp. 267−279.
Change analysis
Tree coverCoastal plain, Bertie county, NC,
N 36.1344° W 77.1074°
no changeincrease in tree coverdecrease in tree cover
Bare groundDallas – Fort Worth, TX, N 32.8694° W 97.0481°
no changeincrease in bare grounddecrease in bare ground
IntroductionSystematic large area land cover monitoring at medium spatial resolution has to date been limited primarily by data costs and processing limitations. However, since January 2008, radiometrically and terrain corrected Landsat data have been made available free of charge over the internet by the USGS Center for Earth Resources Observation and Science (EROS) (Woodcock et al. 2008), and the cost of computer processors and data storage continues to decline. A NASA-funded project, entitled Web-Enabled Landsat Data (WELD), has demonstrated recently the potential of large volume Landsat 7 ETM+ data processing to provide 30m spatially mosaiced and temporally composited timeseries data sets for the conterminous United States (CONUS) and Alaska (http://landsat.usgs.gov/WELD.php) (Roy et al. 2010). This project presents the application of the MODIS (Moderate Resolution Imaging Spectroradiometer) Vegetation Continuous Fields (VCF) characterization approach (Hansen et al. 2003) to the WELD data to produce 30m annual CONUS land cover estimates of percent tree, percent other vegetation (including grasses, shrubs and forbs) and percent bare ground cover, and also the probability of water presence and percent of snow and ice.
MethodsTree-based classification methods are well-established in the remote sensing community and have been used to generate the MODIS Vegetation Continuous Field and Land Cover products (Hansen et al. 2003; Friedl et al. 2002). In this research, a bagged decision tree methodology was used to classify the WELD data. Decision trees are hierarchical classifiers that predict class membership by recursively partitioning a data set into more homogeneous subsets, referred to as nodes (Breiman et al. 1984; Breiman 1996). Trees can accept either categorical data in performing classifications (classification trees) or continuous data in performing sub-pixel percent cover estimations (regression trees). For regression trees, a sum of squares criterion is used to split the data into successively less varying subsets. The split that yields the maximum reduction in variability is selected to build the regression tree model. For annual estimates of percent tree cover and bare ground, regression tree models were employed using a bagging methodology. For preliminary weekly percent water and snow/ice, we applied a linear mixture model, as sub-pixel percent training data were not available. The method is that of Roberts et al. (1998), which employs multiple end-members to linearly model percent cover per pixel. These estimates can then be used as training data for regression tree model implementation.
Deriving percent cover training data using high spatial resolution imagery Visual crown/no crown classifications of thirty-two 4 m Ikonos and three 2.8 m QuickBird multi-spectral images distributed across the CONUS were used to derive sub-Landsat pixel percent crown cover data based on the method described by Hansen et al. (2002) Percent crown cover per 30m WELD grid cell was then calculated and used as training data. Secondly, an additional 3,295,525 30m pixels of either 100% or 0% canopy cover were interpreted visually using very high spatial resolution (<10 m) Google Earth imagery. These additional training pixels were distributed in 2,156 homogeneous polygons across the CONUS and were checked by comparison with the 2008 WELD data to ensure that there was no change between 2008 and the acquisition of the GoogleEarth imagery. Visual interpretation of Google Earth imagery was performed on an ad hoc basis to fill regional gaps where very high spatial resolution Ikonos and QuickBird imagery were lacking.
X – a pixel of mixed class
100*'2'1
'1%dddX
Flowchart of classification process
Metrics
Sampling (25 samples)
Building models (25 regression trees)
Georanking
Median (rank 13) as classification result
Training data
WELD hdf web storage ftp://weldftp.cr.usgs.gov Auxiliary datasets
Classification (25 preliminary classification results)
Deriving training using linear spectral unmixingSpectral Mixture Analysis was used to create training data for snow/ice and water layers. Nearest neighbors from pure classes (0% and 100%) were used as end-members to calculate percent snow/ice and water for pixels of mixed class (1..99%) (Roberts et al., 1998). Euclidean distance measure of closeness was used for this calculation.
Band 5
Band 4
Band 3
Multiyear growing season composite image
trees
other vegetation
bare ground
water
2008 VCF layers as r-g-b composite
Snow and Ice coverCroplands near to Hermiston, OR,
N 45.9012°, W 119.368°Winter mosaic. Bands 5,4,3 as RGB
0% % of snow and Ice cover 100% 0% % of open water 100%
Open water surfaceMarshes close to Sea Breeze, NJ,
N 39.3071° W 75.2616°Summer mosaic. Bands 5,4,3 as RGB
Weekly updated layers
X
Y Z
Y – nearest train pixel of 100% pure class
d2 – distance from X to Y
Direct line connecting 2 and-members (Y and Z)
d1’d2’
d1’ and d2’ are the projections of d1 and d2
Spectral space
Z – nearest train pixel of 0 % pure class
90% percentile75% percentile50% percentile25% percentile10% percentile
Band2 TOA ReflectanceBand3 TOA Reflectance
Band7 TOA ReflectanceBand62 TOA Brightness temperatureBand61 TOA Brightness temperature
Band5 TOA ReflectanceBand4 TOA Reflectance
Band1 TOA Reflectance
d1
Minimum bare ground within the growing season
Tree canopy density within the growing season
New York, NY, NJ, N 40.7628°, W 73.9825°Multiyear composite. Bands 5,4,3 as RGB.
0% % of bare ground 100%
Mount Rainier, WA, N 46.8519°, W 121.774°Multiyear composite. Bands 5,4,3 as RGB.
0% % of tree cover 100%
2006
2007
2008
2009
2010
2006
2007
2008
2009
2010
Annually updated layers
Tree cover > 10,000,000 training pixels derived mostly from crown/no crown classifications of IKONOS and QuickBird imagesBare ground > 150,000,000 training pixels from NLCD impervious surface and MODIS VCF layersWater > 130,000 visually interpreted training pixels with following spectral unmixingSnow/ice > 50,000 visually interpreted training pixels with following spectral unmixing
Tree cover IKONOS and QuickBird images, Google EarthBare ground MODIS VCF, SRTM water bodies and NLCD impervious surface layersWater Google EarthSnow/ice Google Earth
Alexey V. Egorov, Matthew C. Hansen, Peter Potapov, Svetlana Turubanova, David P. Roy, Junchang Ju - South Dakota State University, Brookings, SD NASA MEASURE program grant number NNX08AL93A
d1 – distance from X to Z
d2
100% 1…99% 0%
0% 100%
0% 100%
0% 100%
