Printer Friendly

Classification of pronghorn fawning habitat using Landsat thematic mapper data.

Abstract. -- Management of wildlife resources requires accurate information about habitat characteristics. Six maps of pronghorn (Antilocapra americana) fawning habitat from the Trans-Pecos region of Texas were developed from Landsat 5 Thematic Mapper (TM) data. The goal was to classify fawning habitat and compare unsupervised and supervised habitat classification techniques. Unsupervised and supervised classification training sets were developed for each of two TM scenes taken during the 1990 and 1991 pronghorn breeding season. Data from 12 reference microhabitat variables were used to assess accuracy of the classifications. Additionally, principal components analysis was evaluated as a data reduction technique for the TM data sets. Unsupervised classification accuracy for 1990 and 1991 maps were 41 and 44%, respectively; supervised classification accuracy for each respective map was 29 and 39%. Accuracy of maps after principal components analysis was 31% for unsupervised and 69% for supervised classifications. Supervised training sets performed better than unsupervised training sets in identifying specific fawning habitat signatures. Of the training sets examined, the 1990 supervised set performed the best in identifying critical fawning habitat during drought conditions.

**********

Satellite acquired data sets are useful for mapping large areas of wildlife habitat (Tueller 1987). Landsat Multispectral Scanner (MSS) data have been used to classify elk (Cervus elaphus) habitat in the Blue Mountains of Oregon (Isaacson et al. 1982), American kestrel (Falco sparverius) nesting habitat in Oregon (Lyon 1983), and white-tailed deer (Odocoileus virginianus) habitat in the Saginaw River Basin of Michigan (Ormsby et al. 1985).

Habitat used by pronghorn (Antilocapra americana) during fawning is important to fawn survival (Autenrieth 1982). The ability to remotely identify and assess fawning habitat would be beneficial to pronghorn management programs. Since fawns often occur in specific vegetation types (Pyrah 1974; Autenrieth 1976; 1982; Barrett 1981), vegetation occurring in fawning habitat should produce different spectral signatures, which could be classified from Landsat Thematic Mapper (TM) data.

Although the Trans-Pecos region of Texas represents important habitat for pronghorn, little specific information is available on fawning habitat (Canon 1993). The present study was initiated to evaluate Landsat TM digital data in classifying microhabitat features of pronghorn fawning habitat in the Trans-Pecos region of Texas. Specifically, the objectives were to: (1) evaluate the effectiveness and accuracy of supervised and unsupervised training techniques in a classification scheme using Landsat TM data, and (2) determine if Landsat TM data can be used to classify vegetation used by pronghorn for fawning habitat.

STUDY SITE AND METHODS

The study was conducted in the Trans-Pecos region of Texas, on approximately 38,955 ha of the Double U Ranch in Hudspeth County, Texas. The Double U Ranch is part of the University of Texas Lands System. This area includes some of the best pronghorn habitat in the region (Huber 1992).

The climate is arid; annual precipitation averages 22.5 cm at Cornudas, Texas (NESDIS 1991), which is located 16 km east of the eastern boundary of the study area. Topography is characterized as undulating with broad flat draws (Buechner 1950). Four range sites found on the study area are loamy sites, gravelly sites, draws, and limestone hills and mountains.

Major vegetation types are creosote bush (Larrea tridentata) and tarbush (Florensia cernua) desert shrub, grama (Bouteloua sp.) grassland, and yucca (Yucca sp.) savannah (Correll & Johnston 1979). Dominant vegetation is blue grama (Bouteloua gracilis), burrograss (Scheropogon brevifolius) and tobosagrass (Hilaria mutica), with some areas having considerable densities of succulent and woody plants including yucca (Yucca elata), creosote bush, tarbush, lotebush (Ziziphus obusifolia), allthorn (Koeberlina spinosa), lechuguilla (Agave lechuguilla), and javelina bush (Condalia ericoides).

Landsat 5 TM scenes of the study area were taken on 19 June 1990 and 16 April 1991 to coincide with pronghorn parturition (April-June). Cloud cover on the study area prevented use of same day and month scenes for 1990 and 1991.

Landsat TM data for the 1990 scene were extracted from a magnetic tape in four 1024 by 1024 pixel segments six bands deep. Data were then partitioned (SUBSET; Earth Resources Data Analysis System [ERDAS] 1990) to complete a six band 1024 by 1024 pixel image of the study area that represented 30 by 30 m ground resolution. Problems with downloading band 6, the thermal infrared band, prevented its use in the classification process. The 1991 data were extracted in two 1024 by 1024 pixel segments six bands deep.

Two clustering techniques (STATCL and ISODATA; ERDAS 1990) were used to develop unsupervised training sets for both images. Since it was unknown how spectrally distinct the fawning habitat was, training sets were developed with the maximum number of distinct signatures. A training set with 10 signatures was selected following evaluation of numerous training sets from both clustering techniques (Huber 1992). Ellipse plots (ELLIPSE; ERDAS 1990) were used to evaluate signature distinctness.

Seed clusters were used to build the supervised training sets for both images (SEED; ERDAS 1990). This method allowed the building of training sets using known pixel locations and previous knowledge of pertinent environmental variables. For the 1990 image, characteristics of 14 known fawn bed sites obtained from radio telemetry (Canon 1993) were used to build a signature for the fawning habitat. Signature 10 was the fawning habitat signature for the 1990 error classification matrix of the supervised training set, whereas signature 1 was the fawning habitat signature for the 1991 supervised training set.

The 1991 image was rectified to the 1990 image using nearest-neighbor resampling. Principal components analysis was then performed on the resulting 12 (6 band image for 1990 and 6 band image for 1991) band file (PRINCE; ERDAS 1990). This analysis indicated that a 3-band principal component image was appropriate to capture most (93%) of the variance in the data. The 3-band file was then used to develop unsupervised and supervised training sets. Each training set contained 10 signatures. Signature 1 represented fawning habitat for the principal components supervised training set.

Following development of the training sets, each Landsat data set was classified twice (once for each training set) using a maximum likelihood decision rule (MAXCLAS; ERDAS 1990). A first-pass, parallelepiped classification was used to shorten the time required to run the classification and to avoid potential normality problems of the data (Campbell 1987). These data sets were then stored as Geographic Information System (GIS) files.

The 1990 Landsat data was georectified to the Universal Transverse Mercator grid (GCP, COORDN, and LRECTIFY; ERDAS 1990). A GIS data set was then developed for both of the 1990 classified GIS files. For accuracy assessment of locations, 512 stratified random samples were selected with replacement from each of the two 1990 GIS files. Each sampling location was a 3 by 3 pixel array, representing 90 by 90m. Because sampling was random, the 3 by 3 pixel arrays could include more than one signature type. Consequently, the 200 best locations (also stratified), based on purity of the signature class within the 3 by 3 array, were selected from each set of sampling locations for a total of 400 locations. The locations were then plotted on U.S. Geologic Survey 7.5 minute topographic maps.

All signatures were included in the error classification matrix to show misclassifications to the other signatures. Correctly classified pixels were presented on the diagonal, whereas misclassified pixels were on the off-diagonal. Statistics were computed for those signatures that had sufficient reference data.

Reference samples were collected in the summers of 1991 and 1992. Sampling locations were randomly selected. Twelve reference variables were measured at each location. Ten randomly located 0.1[m.sup.2] foliar cover plots (Hays et al. 1981) were sampled within a 30 m radius of the center of each sampling location. Percent foliar cover (grass, forb and total foliar cover) and rock cover were estimated in each plot. Also, 10 ten-point frames (Hays et al. 1981) were randomly located from which rock, bare ground, vegetation litter and basal cover (grass, forb and total) were recorded. Succulent and woody plant cover was estimated along four 30 by 2 m belt transects (one in each cardinal direction from the center of each sampling location) in which succulent and woody plant canopy and height were measured.

Data from 256 field sample locations were used to assess map accuracy. Variables were standardized prior to clustering. The EML clustering algorithm was used to group the data into meaningful clusters (SAS Institute Inc. 1989). Final clusters contained five variables; these were succulent and woody plant canopy, bare ground, vegetation litter, rock, and total basal cover. Descriptive statistics were calculated for each cluster and used in the accuracy assessment (CLASERR; ERDAS 1990).

Statistics reported for classification error matrices included overall accuracy (total pixels correctly classified/total pixels), user's accuracy (total pixels correctly classified in a signature/total signature locations referenced), producer's accuracy (total pixels correctly classified/total pixels classified to signature). A kappa-like statistic (Foody 1992) was calculated to exclude the possibility of chance agreement in overall accuracy (overall accuracy - 1/n / 1-1/n where n = number of signatures of interest). Clusters were given the same number as the signature they represented in all error matrices.

A GIS file of known fawn locations was constructed (DIGSCRN and GRDPOL; ERDAS 1990), which determined the GIS file values for the classified images (INQUIRE; ERDAS 1990). This data set contained 480 fawn locations, of which 209 were bed-site locations.

RESULTS

Cluster analyses. -- The classification error matrix for the 1990 unsupervised classification included 10 training set signatures and 244 reference locations (Table 1). Only three signatures were referenced sufficiently for inclusion in the error matrix. These signatures received 87% of the pixels classified. Overall classification accuracy was 41%, but the kappa-like statistic was 12%. Clusters 1, 2 and 3 in the 1990 unsupervised classification were primarily draw sites, limestone hills and mountains, and loamy sites, respectively (Table 2).

The error matrix for the 1990 supervised classification included seven reference clusters for signatures 2, 3, 4, 9, 10, 11 and 14, and 231 reference locations (Table 1). These seven signatures received 80% of the pixels in the classification. Signature 10, the fawning habitat signature, had an user accuracy of 37%. Overall accuracy was 26%; the kappa-like statistic generated an accuracy estimate of 14%. Reference clusters 2 and 3 were equated to draw sites, clusters 4 and 14 gravelly sites, cluster 9 limestone hills and mountains, and clusters 10 and 11 loamy sites (Table 2).

The classification error matrix for the 1991 unsupervised classification included four clusters and 235 reference locations (Table 3). The overall accuracy of this map was 44%; user accuracy ranged from 78% for signature 2 to 0% for signature 9 (Table 3). The kappa-like statistic produced an accuracy value of 25%. Cluster 1 represented limestone hills and mountains, cluster 2 loamy sites, cluster 3 draws, and cluster 9 gravelly sites (Table 2). The signatures that were evaluated (1, 2, 3 and 9), received 86% of the pixels in the classification.

The error matrix of the 1991 supervised classification contained five clusters and 230 reference locations (Table 3). Signatures 1 through 5 were sufficiently ground-truthed for accuracy evaluations. Signature 1 was the fawning habitat signature and its accuracy was 40% (user accuracy). Signature 5 also performed well, with user accuracy of 70%. Overall accuracy was 39%; the kappa-like statistic was 23%. Clusters 1 and 3 were equivalent to loamy sites, cluster 2 draws, cluster 4 gravelly sites, and cluster 5 limestone hills and mountains (Table 2).

Principal components analysis. -- Principal components analysis indicated that six bands (1-5 and 7) in the 1990 and 1991 classifications were sufficient in capturing the important characteristics of the data. No single band or subset of bands from either image predominated in principal component 1 (Huber 1992). This indicated that each band was explaining some unique variation.

The error matrix for unsupervised principal components classification included four clusters (1, 2, 3 and 10) and 242 reference locations. Signature 3 performed best (62% user accuracy), whereas signature 10 performed worst (3% user accuracy). The four signatures that were evaluated received 86% of the pixels. Overall accuracy was 31%, but the kappa-like statistic was 10%.

The error matrix for the supervised principal components classification was reduced to two components (clusters 1 and 2) because signature 1, the fawning habitat signature, received 67% of the pixels in the classification. The fawning habitat signature (signature 1) performed well, with an user accuracy of 72%. Overall accuracy for the classification was 69%; however, the kappa-like statistic generated an accuracy estimate to 37%.

GIS analysis. -- About 87% of the pixels were classified to signatures 1, 2 and 3 (Table 4) in the 1990 unsupervised training set. About 95% of the fawn locations and bed sites were in signatures 1, 2 and 3. Similar results were also found in the 1991 unsupervised training set (Table 5) and the principal components unsupervised training set (Table 6).

Classification of the 1990 supervised training set resulted in 40% of the pixels being classed in signatures 3, 9 and 10 (Table 4). About 79% of the fawn bed sites and fawn locations occurred in signatures 3, 9 and 10. Signature 10 (fawning habitat signature) received only 18% of the pixels in the classification but accounted for 44% of the fawn locations and 43% of the bed site locations.

The supervised 1991 training set (Table 5) did not perform as well in identifying fawn locations or bed sites as the training set for the 1990 supervised classification. Signature 10 had 3.7% of all pixels, and accounted for only 2.5 and 2.4% of fawn locations and bed sites, respectively (Table 5). Most pixels were assigned to signatures 1, 3 and 5, which together accounted for 72% of the pixels.

The fawning habitat signature (signature 1) in the principal components supervised training set (Table 6) performed in a similar manner to the 1990 supervised signatures 3, 9 and 10. The classification allocated 67% of the pixels to signature 1; whereas, about 84% of the fawn locations and fawn bed sites were found in signature 1 (Table 6).

DISCUSSION

Ormsby et al. (1985) used Landsat TM data incorporated into a GIS to develop a habitat suitability map for white-tailed deer in Michigan, with classification accuracy ranging from 32 to 81% for different cover types. Huber & Casler (1990) reported preliminary results of elk habitat mapping in Colorado using Landsat TM and digital elevation model data, in which the best results produced were from six classifications yielding 63% and 57% accuracy levels. Results from this study indicated relatively low accuracy for the various methods used to assess pronghorn fawning habitat. This may be due to variability in habitats that are hard to measure remotely, the general characteristics of habitat for fawning, the arid nature and variability of the study area, or the assessment methods used.

Difficulties have been encountered in mapping arid communities (Ustin et al. 1986; Tueller 1987). Belward et al. (1988) noted that cover classes should be selected to have spectral homogeneity as well as ecological significance. Because the reflectance from a 30 by 30 m area can have so many constituents (rock, bare ground, shadows, rock type, species composition, elevation, aspect, slope and others), it is not unusual for confusion to exist. In this study, training set signatures, particularly the unsupervised signatures, were spectrally homogeneous, but in an ecological sense they confuse types that are obviously different. An example would be loamy sites and draws that were confused in every classification. Although reference variables were selected that were used to measure microhabitat characteristics of fawn bed sites described by Canon (1993), those variables could not be used to adequately model or predict brightness levels in the six spectral bands of Landsat data.

The number of signatures per training set may have influenced the results. Possibly too many signatures per training set were developed, which reduced the ecological significance of the signatures by including mixed habitat characteristics. This was apparent in relation to the range of variation found in the reference data. Five or six signatures may have provided a more realistic assessment. However, it was also apparent that too few signatures were developed within the narrow Landsat TM data range surrounding the means in each band, which was evident in the percentage of pixels classified to signatures in the classifications. Each classification, except the 1990 supervised classification, had three signatures that received 70 to 86% of the pixels classified. A more realistic approach may have been to define four or five signatures around the band means with a broad signature in the bright range and a broad signature in the dark range.

Curran & Williamson (1985) found that ground truth data can be less accurate than remotely sensed data. Apparently, in this study, the reference data clusters were a source of error. Also, the variables that were measured probably did not model all of the information contained in the Landsat TM data. Differences in the scale of data collection may have influenced the results since comparisons were made between 30 m circular plots and 3 by 3 pixel arrays (90 by 90 m).

Unfortunately, a classification error matrix could not be constructed from GIS data. There was no way of determining when an area that was not habitat was classified into the habitat signature. If the locations that were referenced for the cluster analysis had been monitored for fawn presence or absence through the study period, then presence or absence at the locations could have been used to construct an error matrix (Congalton 1991).

Fawn locations and bed sites found in a signature in the unsupervised classifications were approximately proportional to the percentage of the pixels classified to a signature. The same was true for the 1991 supervised classification. The result of this proportional indifference was that as a larger percentage of the fawn locations or bed sites were mapped, the same percentage of the study area was also mapped. For instance, about 95% of the fawn locations and bed sites in the 1990 unsupervised classification were in signatures 1, 2 and 3. These signatures included about 86% of the pixels classified (or area mapped). If this represents a fawning habitat map that was 95% accurate, then a blank sheet of paper said to be the fawning habitat map would only be wrong 18% {[1 - (0.86 X 0.95)] X 100} of the time. Such a fawning habitat map would be neither practical nor useful.

Supervised training sets appeared to be more effective in classifying pixels to the fawning habitat signature. In the classification of the principal components supervised training set, the fawning habitat signature received about 67% of the pixels. About 84% of the fawn locations and bed sites were in the fawning habitat signature. However, this situation could have been the result of chance agreement. The fawning habitat signature for the 1990 supervised classification (signature 10) received about 18% of the pixels and 44% of the fawn locations and bed sites. It is unlikely that this is a result of chance agreement alone. From the substantial telemetry data obtained on fawn locations and their corresponding microhabitats (Canon 1993), the 1990 supervised training set was the most effective in mapping pronghorn fawning habitat.

Drought conditions in 1990 limited the area used by pronghorn fawns resulting in the fawning habitat signature receiving only about 18% of the pixels classified. As conditions continually improved through 1991, use of the habitat increased. This resulted in the fawning habitat signature for 1991 receiving about 45% of the pixels classified. The fawning habitat signature in the principal components training set was even higher at 67%. Thus, as the size of the habitat used for fawning increased, so did the area included in the classified signatures. These results suggest that the habitat used by pronghorn in 1990 could be viewed as critical fawning habitat due to the drought conditions that year. A critical fawning habitat map would be more practical to develop than attempting to develop fawning habitat signatures when fawn locations become proportionally indifferent as range conditions return to normal.

CONCLUSIONS

Several conclusions can be drawn from the results of this study. First, a supervised classification using known locations performed better than an unsupervised classification for identifying pronghorn fawning habitat. The GIS analysis indicated that the natural variability in the Landsat data, which the unsupervised clustering algorithm accents, may not effectively model the spectral response of pronghorn fawning habitat. Second, the microhabitat variables that were measured, though possibly important in pronghorn fawning habitat, were incomplete in evaluating spectral response and map accuracy. A multiple regression study to determine the most important variables related to spectral response would be useful to better identify pertinent habitat variables in pronghorn fawning habitat. Third, if the appropriate field reference parameters had been measured for use in evaluating the 1990 supervised classification, the results may have indicated a fairly useful fawning habitat map for drought years. Fawning habitat in years with normal or above normal precipitation appears to be so widespread as to make mapping impractical.
Table 1. Classification error matrix for the 1990 unsupervised and
supervised classifications.

 Unsupervised
Training Reference clusters
set User
signature 1 2 3 Total Acc.

 1 16 1 19 36 44%
 2 12 14 56 82 17%
 3 19 9 69 97 71%
 4 1 6 9 16
 5 0 1 0 1
 6 5 1 2 8
 9 0 2 1 3
10 0 0 1 1
Total 53 34 157 244
Producer
 Acc. 30% 41% 44%

 Supervised
Training Reference clusters
set User
signature 2 3 4 9 10 11 14 Total Acc.

 1 0 0 0 0 1 0 0 1
 2 10 0 1 0 4 10 0 25 40%
 3 7 1 1 0 4 4 0 17 6%
 4 2 0 1 0 7 7 0 17 6%
 5 4 0 0 0 0 0 0 4
 6 0 2 1 5 0 1 0 9
 8 0 0 0 0 7 1 1 10
 9 4 1 1 15 2 5 6 34 44%
10 4 2 0 1 14 14 3 38 37%
11 4 0 0 0 12 17 3 36 47%
12 0 0 0 2 0 1 0 3
14 5 0 3 3 16 8 2 37 5%
Total 40 6 8 27 67 68 15 231
Producer
 Acc. 25% 17% 13% 56% 21% 25% 13%

Training set signature 10 in the supervised classification matrix is the
fawning habitat signature.

Table 2. Mean values of the five habitat variables used to assess the
accuracy of the 1990 and 1991 unsupervised and supervised
classifications.

 1990
 Unsupervised Supervised
 Cluster Cluster
Variable 1 2 3 2 3 4 9 10 11 14

Brush canopy 1.8 3.7 1.4 1.5 2.3 3.3 3.7 1.5 1.0 2.2
Bare ground 30.4 17.3 53.3 33.4 12.6 28.3 17.3 46.4 59.9 45.7
Litter 53.6 21.7 31.9 53.8 75.7 39.5 21.7 38.5 27.4 18.6
Rock 5.3 55.1 8.0 2.7 1.0 18.4 55.1 7.1 5.2 28.1
Basal cover 9.9 5.8 7.6 10.0 10.7 9.7 5.8 7.7 7.5 7.5

 1991
 Unsupervised Supervised
 Cluster Cluster
Variable 1 2 3 9 1 2 3 4 5

Brush canopy 3.7 1.3 1.8 2.2 1.5 1.9 1.0 2.2 3.7
Bare ground 17.3 52.9 30.4 45.7 46.4 30.0 59.9 45.7 17.3
Litter 21.7 33.1 53.7 18.6 38.5 54.0 27.4 18.6 21.7
Rock 55.1 6.2 5.3 28.1 7.1 5.2 5.2 28.1 55.1
Basal cover 5.8 7.6 9.9 7.5 7.7 10.0 7.5 7.5 5.8

Table 3. Classification error matrix for the 1991 unsupervised and
supervised classifications.

 Unsupervised
Training Reference clusters
set User
signature 1 2 3 9 Total Acc.

 1 8 22 4 4 38 21%
 2 1 78 14 7 100 78%
 3 2 32 17 1 52 33%
 4 0 5 7 0 12
 6 5 0 0 1 6
 7 0 2 1 0 3
 8 0 2 2 0 4
 9 10 4 2 0 16 0%
10 1 2 0 0 3
Total 28 147 47 13 235
Prod. Acc. 29% 53% 36% 0%

 Supervised
Training Reference clusters
set User
signature 1 2 3 4 5 Total Acc.

 1 46 19 41 7 1 114 40%
 2 2 6 5 0 0 13 46%
 3 8 15 18 0 1 42 42%
 4 10 2 1 3 8 24 13%
 5 0 2 2 3 16 23 70%
 6 1 3 2 0 0 6
 7 1 0 0 0 0 1
 9 0 2 0 0 0 2
10 3 1 1 0 0 5
Total 71 50 70 13 26 230
Prod. Acc. 65% 12% 26% 23% 62%

Training set signature 1 in the supervised classification matrix is the
fawning habitat signature.

Table 4. Percentage of fawn locations (n = 480) and fawn bed sites (n =
209) in a signature in relation to the percentage of pixels that were
classified to a signature for the 1990 unsupervised and supervised
classifications.

 Unsupervised Supervised
 Fawn Fawn bed Fawn Fawn bed
Signature Pixels locations sites Pixels locations sites
 % % % % % %

 1 20.9 16.7 17.8 0.7 1.0 0.5
 2 29.8 32.3 32.2 8.8 1.5 1.5
 3 36.0 45.8 45.1 12.7 19.6 19.2
 4 6.1 4.6 4.8 6.9 3.8 1.9
 5 0.4 0.0 0.0 4.8 5.4 7.3
 6 3.5 0.2 0.0 6.1 2.1 1.4
 7 0.2 0.0 0.0 1.0 0.2 0.5
 8 0.7 0.4 0.1 5.2 3.1 3.8
 9 1.8 0.0 0.0 10.1 16.3 16.9
10 0.6 0.0 0.0 17.5 44.0 43.1
11 10.8 0.3 1.0
12 1.2 0.2 0.5
13 0.8 0.0 0.0
14 13.4 2.5 2.4
Total 100.0 100.0 100.0 100.0 100.0 100.0

Table 5. Percentage of fawn locations (n = 480) and fawn bed sites (n =
209) in a signature in relation to the percentage of pixels that were
classified to a signature for the 1991 unsupervised and supervised
classifications.

 Unsupervised Supervised
 Fawn Fawn bed Fawn Fawn bed
Signature Pixels locations sites Pixels locations sites
 % % % % % %

 1 14.3 13.1 14.5 44.8 48.9 52.7
 2 37.4 38.2 40.8 7.2 9.2 9.7
 3 24.0 28.6 27.4 14.7 15.6 12.1
 4 4.8 7.7 6.8 9.0 7.5 7.8
 5 0.6 0.0 0.0 13.0 9.6 11.0
 6 3.4 2.3 1.9 4.4 4.4 2.9
 7 0.7 0.8 0.0 1.2 0.4 0.0
 8 2.5 1.0 0.0 1.1 1.9 1.4
 9 9.9 6.0 5.3 0.9 0.0 0.0
10 2.4 2.3 3.3 3.7 2.5 2.4
Total 100.0 100.0 100.0 100.0 100.0 100.0

Table 6. Percentage of fawn locations (n = 480) and fawn bed sites (n =
209) in a signature in relation to the percentage of pixels that were
classified to a signature for the principal components unsupervised and
supervised classifications.

 Unsupervised Supervised
 Fawn Fawn bed Fawn Fawn bed
Signature Pixels locations sites Pixels locations sites
 % % % % % %

 1 20.4 19.4 17.8 67.3 84.2 83.7
 2 16.3 19.6 18.8 2.7 1.0 1.0
 3 36.4 46.2 45.6 1.6 0.0 0.0
 4 0.1 0.2 0.0 6.5 2.9 3.8
 5 6.5 1.7 1.4 6.4 3.8 1.9
 6 0.8 0.4 0.0 4.0 1.2 1.9
 7 0.6 0.0 0.0 1.8 1.5 1.4
 8 6.2 2.9 4.8 4.2 1.0 1.4
 9 0.1 0.0 0.0 0.3 0.6 1.1
10 12.6 9.6 11.6 5.2 3.8 3.8
Total 100.0 100.0 100.0 100.0 100.0 100.0


ACKNOWLEDGMENTS

We thank S. Hartman of the University of Texas Lands for access on the Double U Ranch. We thank S. Meek, who provided field assistance. We are especially indebted to Dr. Ernest B. Fish who provided much of the background support and expertise to see this study through to completion.

LITERATURE CITED

Autenrieth, R. E. 1976. A study of birth sites selected by pronghorn does and the bed sites of fawns. Pronghorn Antelope Workshop Proc., 7:127-134.

Autenrieth, R. E. 1982. Pronghorn fawn habitat use and vulnerability to predation. Pronghorn Antelope Workshop Proc., 10:112-131.

Barrett, M. W. 1981. Environmental characteristics and functional significance of pronghorn fawn kidding sites in Alberta. J. Wildl. Manage., 45:120-131.

Belward, A. S., J. C. Taylor, M. J. Stuttard, E. Bignal, J. Mathews & D. Curtis. 1988. The classification of semi-natural vegetation from Landsat thematic mapper imagery: a users perspective. Proceedings of IGARSS 1988 Symposium, Edinburg, Scotland. 13-16 Sept., 621-622 pp.

Buechner, H. K. 1950. Life history, ecology, and range use of the pronghorn antelope in Trans-Pecos Texas. Am. Midl. Nat., 43:257-354.

Campbell, J. B. 1987. Introduction to remote sensing. Guilford Press, New York, 551 pp.

Canon, S. K. 1993. Fawn survival and bed-site characteristics of Trans-Pecos pronghorn. Unpublished Ph.D. dissertation, Texas Tech Univ., Lubbock, 133 pp.

Congalton, R. G. 1991. A review of assessing the accuracy of classifications of remotely sensed Data. Remote Sens. Environ., 37:35-46.

Correll, D. S. & M. C. Johnston. 1979. Manual of the vascular plants of Texas. Univ. Texas, Richardson, 1881 pp.

Curran, P. J. & H. D. Williamson. 1985. The accuracy of ground data used in remote sensing investigations. Int. J. Remote Sens., 6:1637-1651.

Earth Resources Data Analysis System Inc. 1990. Atlanta, GA.

Foody, G. M. 1992. On the compensation for chance agreement in image classification accuracy assessment. Photogram. Eng. Remote Sensing, 58:1459-1460.

Hays, R. L., C. Summers & W. Seitz. 1981. Estimating wildlife habitat variables. U.S. Fish and Wildlife Service, FWS/OBS-81/47, Washington, D.C., 111 pp.

Huber, G. E. 1992. Classification of pronghorn fawning habitat using Landsat thematic mapper data. Unpublished M.S. thesis, Texas Tech Univ., Lubbock, 92 pp.

Huber, T. P. & K. E. Casler. 1990. Initial analysis of Landsat TM data for elk habitat mapping. Int. J. Remote Sens., 11:907-912.

Isaacson, D. L., D. A. Lechenby & C. J. Alexander. 1982. The use of large-scale aerial photography for interpreting Landsat digital data in an elk habitat-analysis project. J. Appl. Photogr. Eng., 8:51-57.

Lyon, J. G. 1983. Landsat-derived land-cover classifications for locating potential kestrel nesting habitat. Photogram. Eng. Remote Sensing, 49:245-250.

NESDIS, 1991. Climatological data annual summary Texas. NOAA-National Environmental Satellite Information Service.

Ormsby, J. P., J. C. Gervin, R. S. Lunetta & J. Nickeson. 1985. Habitat evaluation and landcover analysis using Landsat-4 TM data. Int. Symp. Remote Sensing Environ. Proc., 19:415-421.

Pyrah, D. 1974. Antelope fawn bedding cover selection in central Montana. Proc. Antelope States Workshop, 6:113-121.

SAS Institute Inc. 1989. SAS/STAT user's guide version 6. SAS Institute, Inc., Cary, North Carolina, 943 pp.

Tueller, P. T. 1987. Remote sensing science applications in arid environments. Remote Sens. Environ., 23:143-154.

Ustin, S. L., J. B. Adams, C. D. Elvidge, M. Rejmanek, B. N. Rock, M. O. Smith, R. W. Thomas & R. A. Woodward. 1986. Thematic mapper studies of semiarid shrub communities. BioScience, 36:446-452.

*Fred C. Bryant and Greg E. Huber

Department of Range, Wildlife and Fisheries Management

Texas Tech University, Lubbock, Texas 79409

*Present Address:

Caesar Kleberg Wildlife Research Institute

Texas A & M University-Kingsville, Campus Box 218

Kingsville, Texas 78363

FCB at: f-bryant@tamuk.edu
COPYRIGHT 1998 Texas Academy of Science
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1998 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Bryant, Fred C.; Huber, Greg E.
Publication:The Texas Journal of Science
Geographic Code:1U7TX
Date:Feb 1, 1998
Words:5403
Previous Article:Reexamination of the range for the northern pygmy mouse, Baiomys taylori (rodentia: muridae), in northeastern Texas.
Next Article:Salinity recovery in Texas bays.
Topics:

Terms of use | Privacy policy | Copyright © 2020 Farlex, Inc. | Feedback | For webmasters