| |
MAPPING SAGEBRUSH/GRASSLANDS FROM LANDSAT TM-7 IMAGERY:
A COMPARISON OF METHODS
INTRODUCTION
We compared three different remote sensing techniques for mapping rangeland vegetation, and sagebrush in particular, from Landsat TM imagery. The first two involved supervised classifications of raster polygons or regions, derived from two different image segmentation methods. In the first method (referred to as SILC3), the original TM-7
imagery (30 m2 pixels) was segmented into raster polygons using a proprietary object and rule-based merging algorithm. In the second method (referred to as eCognition), the TM-7 imagery was resampled to 15 m2, and then segmented into raster polygons using eCognition – a commercially available software product. Both resulting image segmentations then were classified to land cover type, and canopy closure class for sagebrush, using the same training data and classifier(s). The third method (referred to as PCA) was developed by Heather McClure and David Prevedel at the USDA Forest Service, Intermountain Regional Office in Ogden, UT. It involved an unsupervised classification of pixels that had been reassigned spectral values based on a 3-band, principal components image (McClure and Prevedel 2002). The primary objective of our study was to compare the results from the different methods, and in the process to determine which one(s) could, with reasonable accuracy, map sagebrush and greasewood (Atriplex spp.) shrub types according to several different canopy closure classes. A secondary objective was to evaluate the ability of the different methods to accurately classify and map greasewood, plus four distinct sagebrush species.
The study area comprised a portion of Landsat TM-7 scene Path 37/Row 29 in the Clark’s Fork valley of the Yellowstone River in southern Carbon County, Montana, and northern Park County, Wyoming. The full scene was clipped to this smaller area to allow us to focus on sagebrush/grasslands that are most abundant in the Bighorn Basin and to minimize spectral differences in the imagery due to soil or ecological changes.
A copy of the full report is at: BLM Sage Mapping Comparison PDF
SILC3 AND ECOGNITION IMAGE SEGMENTATIONS
Image segmentations were used in the SILC3 and eCognition classifications to first delineate the imagery into regions or polygons based on pixel spectral homogeneity. The SILC3 process was applied to a July 2000 TM-7 image (30 m pixels); 130 spectral classes were created through an iterative, ISODATA cluster analysis, and each pixel to one of these spectral classes. Segmentation was carried out in a multi-step merge process, using a variable minimum map unit (0.1-2.0 ha), to retain unique features based on the spectral heterogeneity of surrounding pixels. In the eCognition process, a region-growing algorithm was used to segment a September 1999 Landsat TM-7 image resampled to 15m. The eCognition software uses a proprietary algorithm to delineate ground features based on the spectral characteristics of individual pixels and the shape of expanding regions. The eCognition segmentation produced about twice as many regions as SILC3, and did a much better job of capturing sagebrush mosiacs and linear sagebrush features. Regions from both segmentations were attributed with average TM band values from the imagery and topographic values (elevation, aspect, and slope) from a 7.5 minute DEM.
SILC3 and eCognition image segmentations for Dilworth Creek, southwest of Belfry Montana.
The area is a sagebrush mosiac with Wyoming Big Sagebrush and grass, it also has riparian grasslands along Dilworth Creek.
| SILC3 regions on 30m Landsat imagery | eCognition regions on resampled 15m Landsat imagery |
 |  |
| Note smaller region size and linear features in eCognition segmentation, especially along streams. |
SILC3 and eCognition image segmentations for Bear Creek in southern Pryor Mountains.
The area is where Wyoming Big Sagebrush and grasslands border Utah Juniper at the base of the Pryor Mountains.
| SILC3 regions on 30m Landsat imagery | eCognition regions on resampled 15m Landsat imagery |
 |  |
| Again note increase of linear features captured in eCognition segmentation. |
PCA PIXEL CLASSIFICATION
In the PCA analysis, all non-rangeland types except Utah Juniper and Limber Pine were masked out of the resampled (15m pixel) September 1999 Landsat TM-7 imagery using the 1992 National Land Cover Data (NLCD). The masked imagery was then run through a principal components analysis to produce the top three principal component bands. The first three principal component bands were then used in an unsupervised classification where 69 spectral signature classes were created in an iterative, ISODATA cluster analysis. The PCA image pixels were then classified into one of the spectral class groups using the spectral signature. In a final step "Salt and Pepper" effects were removed
using Guided Clustering.
PCA unsupervised classification into spectral classes before region grouping for Dilworth Creek and Bear Creek.
| Dilworth Creek | Bear Creek |
 |  |
| Note linear and mosaic features captured in spectral class pixel classifications. |
LABELING PROCESS
Supervised classifications using the same training data set were used to assign labels to the SILC3 and eCognition image segmentation regions. New training data were collected for all rangeland types and combined with an existing training data set. The training data were evaluated with the imagery for positional accuracy and lifeform agreement, and then attributed with mean spectral and topographic values for the pixels in the region in which each training datum fell. Labels were assigned in a supervised classification using a Euclidian (Dudani) distance-weighted classifier and a mean inverse distance spatial adjustment (Steele et al. in press). In the PCA method, each spectral class was assigned a label based on the majority rangeland type for all training data that fell within all pixels of the spectral class. Twenty PCA spectral classes associated with barren or water lacked any associated training data and were classified manually.
Sagebrush canopy cover maps for Dilworth Creek, southwest of Belfry Montana.
Map Legend: LtGreen-VLow Cover Grass (3130), MedGreen-Low/Mod Cover Grass (3150), DrGreen-Mod/High Cover Grass (3170), Tan-Sagebrush/Xeric Shrub 5-14% (3370), LtBrown-Sagebrush/Xeric Shrub 15-24% (3380), DrBrown-Sagebrush/Xeric Shrub 25-34% (3390), Orange-Mesic Shrub/Willow (3610), Purple-Utah Juniper (4214), the red lines are roads and blue lines are streams.
| SILC3 Classification | eCognition Classification |
 |  |
| PCA Classification |
 |
Sagebrush canopy cover maps with training data for Bear Creek in southern Pryor Mountains.
Map Legend: LtGreen-VLow Cover Grasslands (3130), Tan-Sagebrush/Xeric Shrub 5-14% (3370),
LtBrown-Sagebrush/Xeric Shrub 15-24% (3380), Purple-Utah Juniper (4214), Gray-Rock/Barren (7300), the
red lines are roads and blue lines are streams.
| SILC3 Classification | eCognition Classification |
 |  |
| PCA Classification |
 |
RESULTS
All land cover types within the study area were mapped in the SILC3 and eCognition classifications. In the PCA classification all non-rangeland types except Juniper and Limber Pine were masked out from the classification using the 1992 NLCD land cover classification. All three methods mapped similar amounts of sagebrush/xeric shrubs, which
were almost twice the amount of the MTGAP or NLCD classifications. However, looking at the maps above, each of the methods mapped sagebrush/xeric shrubs in different locations. SILC3 appeared to underclassify grasslands in favor of non-rangeland types, and both SILC3 and eCognition classified more non-rangeland types than the PCA method.
Acreage table for sagebrush canopy cover and rangeland classes
| Covertype Code | NAME | SILC3 | ECOG | PCA |
|---|
| 3130 | Very Low CoverGrasslands | 111,561 | 140,911 | 111,191 |
| 3150 | Low / Moderate Cover Grasslands | 32,006 | 22,437 | 45,231 |
| 3170 | Moderate / High Cover Grasslands | 5,888 | 12,859 | 24,558 |
| 3370 | Sagebrush / Xeric Shrubs 05-14% | 150,555 | 193,472 | 191,480 |
| 3380 | Sagebrush / Xeric Shrubs 15-24% | 206,074 | 160,128 | 170,683 |
| 3390 | Sagebrush / Xeric Shrubs 25-34% | 66,469 | 41,637 | 71,028 |
| 3395 | Sagebrush / Xeric Shrubs >= 35% | 17,500 | 27,292 | 8,254 |
| 3610 | Mesic Shrubs / Willow | 4,578 | 4,592 | 17,222 |
| 4205 | Limber Pine | 3,144 | 4,588 | 4,532 |
| 4216 | Utah Juniper | 56,166 | 53,230 | 34,505 |
| 7300 | Rock / Barren | 32,573 | 24,249 | 38,946 |
CONCLUSIONS
Small or narrow landscape features like woody draws, sagebrush stringers, or patches of mesic shrubs, willows, or aspen appeared to be better captured by the 15m imagery used in the eCognition and PCA methods, than by 30m data used by SILC3. The PCA technique was the fastest method to produce an initial labeled map, albeit for sagebrush/rangeland types only. It required less training data than the other methods but did not have any accuracy estimate. The SILC3 and eCognition methods took longer to produce initial maps due to the amount of training data required, but did have accuracy estimates as part of the process.
All three techniques were able to classify sagebrush canopy classes and sagebrush species types. Although we believe that they represent substantial improvements over existing land cover datasets, such as MTGAP or NLCD, we strongly advocate the need for an independent accuracy assessment of the three map outputs. This should be relatively easy to accomplish and will help objectively determine if one method is substantially better than the others or how these three methods compare with other available ones.
REFERENCES
McClure, H., and D. Prevedel. 2002. Unpublished report. USDA Forest Service, Intermountain Region, Ogden, UT.
Steele, B., Patterson, D., and R. Redmond. in press. Estimating thematic map accuracy without a probability test sample. Ecological and Environmental Statistics.
Sagebrush mapping metadata and Sagebrush Canopy Cover GIS layer for each method
Note: GIS layers are in Erdas Imagine 8.5 format and are in MT-albers projection which is slightly different than the MTBLM-albers projection.
Project completion date: December 12, 2002.
| Principal Investigator | Roland L. Redmond |
| Image/GIS Analysts | Chip Fisher, Will Gustafson |
Funding provided by the Bureau of Land Management, Montana/Dakota State Office, Billings, MT
|
|
