Two image segmentation techniques and a pixel classification were compared for mapping sagebrush in southern Montana. The SILC3 method used an image segmentation based on a 30m unsupervised classification of a July 2000 Landsat image. The eCognition method used an image segmentation based on a 15m bottom up region merging algorithm of a September 1999 Landsat image. The modified USFS Region Four principal components analysis (PCA) method used a pixel classification based on a 15m isodata analysis of the first 3 principle components of a September 1999 Landsat image. A study area was defined for the project based on the intersection of the July 2000 and September 1999 Landsat imagery. To reduce spectral and ecological differences, the study area was additionally restricted to the intersection of the two Landsat images within Montana and 24 miles south of the Montana border in Wyoming. Below is a detailed description of how the BLM3729ECOG grid was created using the eCognition image segmentation method.
BLM3729ECOG Land Cover Type Classification using Sagebrush Canopy Classes Covertype Acres Description ------------------------------------------------ 1100 3937.1 Urban or Developed Lands 2010 82499.4 Agricultural - Dry 2020 171705.7 Agricultural - Irrigated 3130 542839.6 Very Low Cover Grasslands 3150 368958.0 Low / Moderate Cover Grasslands 3170 43581.6 Moderate / High Cover Grasslands 3370 416683.7 Sagebrush / Xeric Shrubs 05-14% Cover 3380 437902.4 Sagebrush / Xeric Shrubs 15-24% Cover 3390 130989.9 Sagebrush / Xeric Shrubs 25-34% Cover 3395 63706.3 Sagebrush / Xeric Shrubs >= 35% Cover 3610 134172.8 Mesic Shrublands / Willow 4101 18069.6 Aspen 4150 32303.1 Mixed Broadleaf / Cottonwood Forest 4203 29811.1 Lodgepole Pine 4204 108338.3 Whitebark Pine 4205 81662.1 Limber Pine 4206 38666.5 Ponderosa Pine 4212 143962.2 Douglas-fir 4216 175274.3 Utah Juniper 4223 5185.4 Douglas-fir/Lodgepole Pine 4237 115601.7 Subalpine Fir/Spruce 4241 74013.2 Mixed Upper Subalpine Conifer Forest 4242 31861.2 Mixed Lower Subalpine Conifer Forest 4244 25927.0 Mixed Xeric Conifer Forest 4400 342704.2 Burns 5000 30281.1 Water 7300 325728.8 Rock/Barren 7500 484.8 Mines/Quarries 9100 35524.5 SnowBLM3729ECOG Sagebrush Species Classes (for regions classified as a sagebrush type above) Covertype Acres Description ------------------------------------------------ 3311 35802.5 Greasewood 3351 28977.4 Mountain Big Sagebrush 3352 861631.5 Wyoming Big Sagebrush 3353 9605.0 Basin Big Sagebrush 3354 113265.9 Black / Low Sagebrush
First, the September 1999 Landsat TM data was clipped to the study area and then resampled to 15m using the panchromatic band. Next an NDVI layer was created using TM bands 3 and 4:
NDVI = (TM4 - TM3) / (TM4 + TM3)
An image segmentation was run in eCognition using a bottom up region growing algorithm. First the TM band and NDVI layers were loaded into the eCognition program. Segmentation settings for scale and color/shape contrast were selected and image segmentations were created. Ground feature delineations were analyzed using DOQ's and the original imagery. Additional segmentation settings for shape features are also used. Lower scale factors resulted in smaller minimum map units. Setting the color/shape contrast to favor pixel color separation verses region shape resulted in more linear feature delineation. This was particularly important in delineating draws with high canopy cover sagebrush. A scale factor of 10 and a color/shape contrast of 0.8 to 0.2 were selected for the final image segmentation. The image segmentation was exported from eCognition and imported into an ARC/INFO region grid. Each region was assigned a set of attributes for TM channels 1-7 and panchromatic from the imagery, along with elevation, aspect, and slope from the 30 meter digital elevation model (DEM); these attribute values were calculated based on the mean values of the 30 meter pixels within each region.
Vegetation overstory (land cover type) was labeled using a supervised classification. The land cover type supervised classification was performed using a combination of manual and training data classifications. The manual classification was performed by on screen analysis of the imagery and directly assigning cover type labels. The labels were filled into the MANLABEL attribute. The following classes were manually labeled:
COVERTYPE DESCRIPTION ------------------------------------------------ 1100 Urban or Developed Lands 2010 Agricultural - Dry 2020 Agricultural - Irrigated 4400 Burns 5000 Water 7500 Mines/Quarries
Twenty One land cover types were labeled using a training data classification. Canopy cover classes were used for both grassland and sagebrush/xeric shrub types because canopy classes provided significantly better separation between the two lifeforms than species classes did. A euclidian distance classifier based on TM band data and topographic variables with a nearest neighborhood of 15 was used in conjunction with a mean inverse spatial classifier to assign labels to the regions. The labels were filled into the COV_CODE_1 ATTRIBUTE. The following classes were used in the training data classification:
COVERTYPE DESCRIPTION ------------------------------------------------ 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 Forest 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 4240 Mixed Conifer Forest 7300 Rock/Barren 9100 Snow
After the training data classification, regions labeled as Mixed Conifer Forest (4240) were assigned to one of three mixed conifer types using a ruleset based on elevation and PVT type for the region (see mixcon_recode.aml).
4241 Mixed Upper Subalpine Conifer Forest 4242 Mixed Lower Subalpine Conifer Forest 4244 Mixed Xeric Conifer Forest
A second training data classification was used to assign sagebrush species to regions classified as sagebrush/xeric shrub types in the first train data classification. A euclidian distance classifier based on TM band data and topographic variables with a nearest neighborhood of 15 was used in conjunction with a mean inverse spatial classifier to assign labels to the regions. The labels were filled into the SAGESPP1 ATTRIBUTE. The following classes were used in the sagebrush species classification:
SAGESPECIES DESCRIPTION ------------------------------------------------ 3311 Greasewood 3351 Mountain Big Sagebrush 3352 Wyoming Big Sagebrush 3353 Basin Big Sagebrush 3354 Black / Low Sagebrush
An accuracy assessment was performed on the land cover type and sagebrush species labels assigned from the training data classifications. A cross validation (leave-one-out) accuracy assessment was performed using the entire training data set.
To help track similarities and differences in map legends between this classification and previous ones (SILC1 and MT-GAP) cover type names and codes numbers were not changed if the actual types were mapped using similar training data.
Users Accuracies of Land Cover Types were: Acc% Covertype Definition ------------------------------------------- 82.5 3130 Very Low Cover Grasslands 81.7 3150 Low / Moderate Cover Grasslands 80.2 3170 Moderate / High Cover Grasslands 56.2 3370 Sagebrush / Xeric Shrubs 05-14% Cover 52.7 3380 Sagebrush / Xeric Shrubs 15-24% Cover 50.5 3390 Sagebrush / Xeric Shrubs 25-34% Cover 84.0 3395 Sagebrush / Xeric Shrubs >= 35% Cover 74.4 3610 Mesic Shrublands / Willow 68.0 4101 Aspen 74.3 4150 Mixed Broadleaf / Cottonwood Forest 70.2 4203 Lodgepole Pine 75.8 4204 Whitebark Pine 67.3 4205 Limber Pine 75.8 4206 Ponderosa Pine 78.1 4212 Douglas-fir 75.7 4216 Utah Juniper 38.4 4223 Douglas-fir/Lodgepole Pine 56.1 4237 Subalpine Fir/Spruce 65.5 4240 Mixed Conifer Forest 99.0 7300 Rock/Barren 100.0 9100 Snow
Users Accuracies of Sagebrush Species were: Acc% SAGESPECIES DESCRIPTION ------------------------------------------------ 76.4 3311 Greasewood 89.8 3351 Mountain Big Sagebrush 97.5 3352 Wyoming Big Sagebrush 33.3 3353 Basin Big Sagebrush 67.8 3354 Black / Low Sagebrush
NDVI = (TM4 - TM3) / (TM4 + TM3)
An image segmentation was run in eCognition using a bottom up region growing algorithm. First the TM band and NDVI layers were loaded into the eCognition program. Segmentation settings for scale and color/shape contrast were selected and image segmentations were created. Ground feature delineations were analyzed using DOQ's and the original imagery. Different weights were assigned to the TM and NDVI bands which had and effect on the color segmentation limits. Additional region shape limits for region smoothness and compactness shape were also used. The eCognition region growing alogrithm attempts to build the most homogenous region possible based in the segmentation settings entered. Lower scale factors resulted in smaller minimum map units. Setting the color/shape contrast to favor pixel color separation verses region shape resulted in more linear feature delineation. This was particularly important in delineating draws with high canopy cover sagebrush. The final image segmentation used the following settings:
Band Weights Scale Color Shape Shape-Additional TM1 TM2 TM3 TM4 TM5 TM7 NDVI (smooth) (compact) --------------------------------------------------------------------- 0.5 0.5 1.0 1.0 1.0 1.0 2.0 10 0.8 0.2 0.9 0.1
The final image segmentation was exported from eCognition and imported into an ARC/INFO region grid. The final region grid had 549,246 regions with an average size of 9.0 acres.
MAKE-HYDROGRAPHY: extracted a hydrography coverage from the WSAL hydrography database for use during labeling.
ATTRIBUTE-GRID: Filled in VALUE, COUNT, HECTARES, PERIMETER, TM1, TM2, TM3, TM4, TM5, TM7, PAN, MNDVI, ELE, SLP, ASP, and SLPASP fields.
PROCESS-GT: filled in spectral, topographic, and hydrologic information in ground-truth training data for use in supervised classification.
CODE DESCRIPTION MNDVI ----------------------------------------------- 3130 very low cover grasslands < -10 3150 low / moderate cover grasslands -10 to 34.9 3170 moderate / high cover grasslands >= 35
Note: Snow was classified using training data. Normally snow is manually classified based on spectral classes associated with snow. However, ecognition does not produce spectral classes, thus snow was classified using training data.
The training data were run through a cross-validation test where each train data was removed and then classified using the remaining train data set. A Dudani distance-weighted euclidian distance classifier using variables: TM1, TM2, TM3, TM4, TM5, TM7, MNDVI, ELEV (scaled elevation), SLP, and SLPASP with a nearest neighborhood of 15 was used in conjunction with a mean inverse spatial classifier. Outlier train data were identified based on the cross-validation test and were removed from the train data set. After completing the train data evaluation, the final train data set was used to run a supervised classification of all regions in the BLM3729ECOG grid. The euclidian distance and spatial classifiers described above were used to classify the regions. The top three most probable land cover types were filled into the COV_CODE_1, COV_CODE_2, and COV_CODE_3 fields.
The final land cover type label was filled into the COVERTYPE field from the MANLABEL field if it was > -1 else from the COV_CODE_1 field.
Regions classified as Mixed Conifer Forest (4240) were recoded to one of three types based on a ruleset using the pvt type and elevation of the region.
4241 Mixed Upper Subalpine Conifer Forest 4242 Mixed Lower Subalpine Conifer Forest 4244 Mixed Xeric Conifer Forest
The aml with the ruleset is available through the data contact.
COVERTYPE DESCRIPTION ------------------------------------------------ 1100 Urban or Developed Lands 2010 Agricultural - Dry 2020 Agricultural - Irrigated 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 Forest 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 Forest 4242 Mixed Lower Subalpine Forest 4244 Mixed Xeric Conifer Forest 4400 Burns 5000 Water 7300 Rock/Barren 9100 Snow
COVERTYPE DESCRIPTION ------------------------------------------------ 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 Forest 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 4240 Mixed Conifer Forest 7300 Rock/Barren 9100 Snow
Code Definition ---------------- -1 Not Manually Classified 1100 Urban or Developed Lands 2010 Agricultural - Dry 2020 Agricultural - Irrigated 4400 Burns 5000 Water 7500 Mines / Quarries
Code Description ----------------- 0 Flat 1 North 2 Northeast 3 East 4 Southeast 5 South 6 Southwest 7 West 8 Northwest
Code Description ----------------- 3311 Greasewood 3351 Mountain Big Sagebrush 3352 Wyoming Big Sagebrush 3353 Basin Big Sagebrush 3354 Black / Low Sagebrush
Code Definition ---------------- 0 no pvt code assigned 1 abla1 (lower elevation subalpine fir type) 2 abla2 (high elevation subalpine fir type) 6 picea (spruce type) 7 pincon (lodgepole pine) 8 pifl (limber pine type) 9 pinpon (ponderosa pine) 11 psemen (Douglas-fir type) 12 agric (agriculture type) 13 agrsmi 14 alpgrass (alpine grasslands type) 15 drygrass (xeric, valley bottom or foothills grassland type) 16 fesida (Idaho fescue type) 18 ripgrass (riparian grass type) 19 rock (rock dominated type) 20 bigsage (big sagebrush type) 21 dryshrub (non-sagebrush xeric shrub type) 22 messhrub 23 potfru (shrubby cinquefoil type) 25 ripshrub (riparian shrub type) 27 urban 33 pial/laly (white bark pine/alpine larch)