First a modified USFS Region Four principal components analysis (PCA) method pixel classification of range land types with sagebrush/xeric shrub canopy cover and species classes was performed. As part of that process non-rangeland areas were masked out of the imagery and were not classified. A sagebrush canopy and land cover type grid was created from the BLM3729R4PCA.covertype attribute.
The non-rangeland areas were filled in from the BLM3729SILC3 classification (see metadata file: BLM3729SILC3.htm). A sagebrush canopy and land cover type grid was created from the BLM3729SILC3.covertype attribute and was then resampled to 15m using Nearest Neighbor resampling. The final sagebrush canopy and land cover type grid was created by merging the r4pcasagecan and silc3sagecan grids.
PCA_S3_SAGCAN Sagebrush Canopy and Range Lands Classes Covertype Acres Description ------------------------------------------------ 1100 2693.6 Urban 2010 42613.5 Dry Agriculture 2020 124090.3 Irragated Agriculture 3130 483737.8 Very Low Cover Grasslands 3150 459408.9 Low / Moderate Cover Grasslands 3170 223612.9 Moderate / High Cover Grasslands 3370 505553.5 Sagebrush / Xeric Shrubs 05-14% Cover 3380 312530.4 Sagebrush / Xeric Shrubs 15-24% Cover 3390 223739.2 Sagebrush / Xeric Shrubs 25-34% Cover 3395 51882.0 Sagebrush / Xeric Shrubs >= 35% Cover 3610 205878.3 Mesic Shrublands / Willow 4101 11577.0 Aspen 4150 18424.1 Mixed Broadleaf/Cottonwood 4203 28635.6 Lodgepole Pine 4204 95240.7 Whitebark Pine 4205 100575.3 Limber Pine 4206 30236.0 Ponderosa Pine 4212 133031.7 Douglas-fir 4216 199067.2 Utah Juniper 5000 4801.2 Water 4223 4845.7 Douglas-fir/Lodgepole Pine 4237 108304.8 Subalpine Fir/Spruce 4241 65112.7 Mixed Upper Subalpine Conifer Forest 4242 28020.9 Mixed Lower Subalpine Conifer Forest 4244 20761.4 Mixed Xeric Conifer Forest 4400 177711.3 Recent Burns 5000 28342.5 Water 7300 299789.5 Rock/Barren 7500 107.9 Mines/Quarries 9100 18768.3 Snow
Modified Region Four PCA Method: First, the September 1999 Landsat TM data was clipped to the study area and then resampled to 15m using the panchromatic band. A mask of non-range lands was created using the 1992 NLCD Land Cover from Montana and Wyoming. The 15m resampled TM data was then clipped to only range land/barren types using the NLCD mask. This was done to reduce spectral differences in the imagery and focus the classification on range land types. The Following NLCD classes were used to create the non-range land mask, the remaining NLCD cover types were not in the mask and were areas to be classified using the September 1999 Landsat imagery.
CODE Status NLCD Cover Description ------------------------------------------ 11 Masked Open Water 12 Masked Perennial Ice/Snow 21 Masked Low Intensity Residential 22 Masked High Intensity Residential 23 Masked Commercial/Industrial/Transportation 32 Masked Quarries/Strip Mines/Gravel Pits 41 Masked Deciduous Forest 42 Masked Evergreen Forest 43 Masked Mixed Forest 81 Masked Pasture/Hay 82 Masked Row Crops 83 Masked Small Grains 84 Masked Fallow 91 Masked Woody Wetlands31 Classified Bare Rock/Sand/Clay 33 Classified Transitional 51 Classified Shrubland 71 Classified Grasslands/Herbaceous 85 Classified Urban/Recreational Grasses 92 Classified Emergent Herbaceous Wetlands
A Principal Components Analysis was run on the clipped and masked image. The six TM band image (1,2,3,4,5,7) was reduced to a three band PCA image. The three band PCA image was then run through an unsupervised classification using an ISODATA clustering algorithm. An initial set of 45 spectral classes and their spectral signatures were created. The resulting image from the unsupervised classification was then examined in relation to a training data set. Five spectral classes were identified with a high level of lifeform confusion based on the training data points. These five confused spectal classes were removed from the original 45 spectral class set, run through a second ISODATA analysis where 10 classes were separated, and the new 10 spectral class set was added back to the remaining original 40 spectal class set. A new supervised classification was run on the imagery and again inspected in relation to the training data set. The process of identifying spectal classes with a high level of lifeform confusion, removal, spliting, and adding back to the spectal class set was done 2 additional times. A final spectral class set of 70 classes was created and a final supervised classification of the imagery was performed. The result was an image with every pixel assigned to one of the seventy spectral classes.
Each of the 70 spectral classes in the final unsupervised classification image was assigned to a sagebrush canopy cover or range land class using training data, DOQ's, and field inspections. Additionally, sagebrush species classes were assigned to spectral classes labeled as a sagebrush canopy class.
The spectral class image was converted to an Arc/Info value grid and then run through a regiongroup function to create a very large region grid. A final region grid with a minimum map unit of 30m (4-15m pixels) was created using the Arc/Info nibble command. The sagebrush and range land classes were added as attributes to the final region grid. Topographic attributes Elevation, Aspect, and Slope were added to the region grid using a DEM grid. Additionally, a potential riparian zone attribute was added to the region grid using a 60m buffer around major streams.
A ruleset based on topographic and geographic limits was developed to recode regions incorrectly classified in the sagebrush and range land classification. For example regions classified as 3353 (Wyoming Big Sagebrush) occuring in the Pryor Mountains above 6500 feet were recoded to 3351 (Mountain Big Sagebrush). The ruleset changes were inspected using training data, DOQ's, and field maps. Adjustments were made to ruleset based on the training data, DOQ's, and field maps. The final sagebrush canopy cover and range land codes were filled into the SAGECANOPY attribute and the final sagebrush species and range land codes were filled into the SAGESPECIES attribute.
No accuracy assessment was performed on either the sagebrush canopy cover or sagebrush species and range land classifications. It is hoped that an independent accuracy assessment of all three sagebrush mapping methods can be completed in the future.
CODE Status NLCD Cover Description ------------------------------------------ 11 Masked Open Water 12 Masked Perennial Ice/Snow 21 Masked Low Intensity Residential 22 Masked High Intensity Residential 23 Masked Commercial/Industrial/Transportation 32 Masked Quarries/Strip Mines/Gravel Pits 41 Masked Deciduous Forest 42 Masked Evergreen Forest 43 Masked Mixed Forest 81 Masked Pasture/Hay 82 Masked Row Crops 83 Masked Small Grains 84 Masked Fallow 91 Masked Woody Wetlands31 Classified Bare Rock/Sand/Clay 33 Classified Transitional 51 Classified Shrubland 71 Classified Grasslands/Herbaceous 85 Classified Urban/Recreational Grasses 92 Classified Emergent Herbaceous Wetlands
A Principal Components Analysis was run on the clipped and masked image. The six TM band image (1,2,3,4,5,7) was reduced to a three band PCA image using the Erdas Imagine principal components function (number of components set to 3, 8-bit stretch, and ignore zero in stats options were selected).
The three band PCA image was then run through an unsupervised classification using an ISODATA clustering algorithm. An initial set of 45 spectral classes and their spectral signatures were created using the Erdas Imagine unsupervised classification function (number of classes - 45, max iterations - 15, convergence threshold - 0.95, and classify zeros was not selected). The resulting image from the unsupervised classification was then examined in relation to a training data set. Five spectral classes were identified with a high level of lifeform confusion based on the training data points. These five confused spectal classes were removed from the original 45 spectral class set (leaving 40 classes), and run through a second unsupervised classification in Imagine (number of classes - 10, max iterations - 6, convergence threshold - 0.95, and classify zeros was not selected). The new spectral class set of 10 signatures was added back to the remaining original 40 spectal class set. A new supervised classification was run on the imagery using the new 50 spectral class set using the Erdas Imagine supervised classification function (non-parametric rule - none, parametric rule - maximum likelihood, and classify zeros not selected). The resulting image was again inspected in relation to the training data set. The process of identifying spectal classes with a high level of lifeform confusion, removal, spliting, and adding back to the spectal class set was done 2 additional times. A final spectral class set of 70 classes was created and a final supervised classification of the imagery was performed. The final result was an image with every pixel assigned to one of the seventy spectral classes.
MAKE-DEM: extracted a digital elevation model (DEM) from the WSAL database for use during labeling.
MAKE-HYDROGRAPHY: extracted a hydrography coverage from the WSAL hydrography database for use during labeling.
COVERTYPE DESCRIPTION ------------------------------------------------ 1100 Urban 2010 Dry Agriculture 2020 Irragated Agriculture 3130 Very Low Cover Grasslands 3150 Low / Moderate Cover Grasslands 3170 Moderate / High Cover Grasslands 3370 Sagebrush / Xeric Shrubs 05-14% Cover 3380 Sagebrush / Xeric Shrubs 15-24% Cover 3390 Sagebrush / Xeric Shrubs 25-34% Cover 3395 Sagebrush / Xeric Shrubs >= 35% Cover 3610 Mesic Shrublands / Willow 4101 Aspen 4150 Mixed Broadleaf / Cottonwood 4203 Lodgepole Pine 4204 Whitebark Pine 4205 Limber Pine 4206 Ponderosa Pine 4212 Douglas-fir 4216 Utah Juniper 4223 Douglas-fir/Lodgepole Pine 4237 Subalpine Fir/Spruce 4241 Mixed Upper Subalpine Conifer Forest 4242 Mixed Lower Subalpine Conifer Forest 4244 Mixed Xeric Conifer Forest 4400 Recent Burns 5000 Water 7300 Rock/Barren 7500 Mines/Quarries 9100 Snow