thresholds within the 67 spectral classes in order to 
reclassify the data into more definitive land cover 
categories. After working with the thermal data it 
was decided to add two more land cover categories to 
the classification. The light vegetation class was 
divided into a class of sparse vegetation with soil 
showing through, and another class of mostly soil 
with a little green vegetation remaining. The mixed 
pixels were divided into three groups instead of two. 
These categories include mixed pixels with a low 
percentage of vegetation (10-40 percent), medium 
vegetation response (around 0.50) , and mixed pixels 
with a dominant vegetation component (over 0.75). 
These two additional categories raised the number of 
land cover types that were identified to 14. 
Fran a thermal channel 6 histogram it was often 
possible to observe "breaks," or places where bimodal 
distribution curves crossed. It was mentioned, for 
example, that class 54 of the classification map 
represented both open water and coal. The mean 
multispectral signatures for these two cover types 
were nearly identical yet the thermal band 
demonstrated a clear distinction between the two 
categories. Other class separations were not as 
clear as in class 54. A more typical situation is 
class 56, where stubble fields were classed the same 
as the mixed cover in several trailer courts. The 
thermal channel indicted that agricultural fields 
were slightly cooler than their urban residential 
counterparts. After changes were made the map was 
again reclassified and grouped into the 14 cover 
types. The reclassified map with 122 classes is 
referred to as THRM67. 
4 ACCURACY EVALUATION OF LAND COVER MAPS 
In order to compare the utility of the two land cover 
maps (THRM67 and CLUS67) an accuracy assessment was 
made on each. The first step was to geometrically 
correct the maps so that each classified pixel 
corresponded to a particular Universal Transverse 
Mercator (UTM) coordinate on the ground. 
A random sample stratified by map categories was 
selected as the procedure for assessing map accuracy. 
An ELAS module named "RANS" (uniform random sample) 
created a new data file from the geometrically 
corrected THRM67 classification map. This file 
contained the original classification data minus 350 
sample locations (25 samples each from the 14 land 
cover categories). A cursor was trained on each 
sample pixel to determine the precise UTM coordinates 
for each sample. These coordinates were then 
plotted onto USGS 1:24,000 topographic maps. When 
all 350 samples were plotted, an investigation 
ccnmenced to determine the "ground truth" category 
for each of the sample sites. In an attempt to 
eliminate the bias, the map categories were not 
included on the maps taken to the field for ground 
investigation. 
When the "ground truth" determination was 
canpleted, the observed cover categories were 
compared to the map categories and a confusion matrix 
was constructed. Percentages were calcuated for the 
number of pixels correctly classified, pixels that 
were classified in the wrong category (errors of 
catmission), and pixels that were not placed in the 
correct categories (errors of omission). The map 
category marginal proportions were calculated in 
order to remove the over-represenation of small 
categories from the stratified randan sample (Card 
1982). These marginal proportions were multiplied by 
the percentages of correctly classified pixels and 
then surtmed to provide the overall map classification 
accuracy (estimated probability correct). 
The same 350 samples derived from the THRM67 
classification were also used to build a confusion 
matrix for the CLUS67 map. The CLUS67 map 
demonstrated an overall map accuracy of 80.2 percent 
for the fourteen land cover categories, with a 0.05 
confidence interval between 0.7581 - 0.8459. The 
THRM67 map, on the other hand, showed an accuracy of 
91.6 percent with a 0.05 confidence interval of 
0.8866 - 0.9434. The overall accuracy assessments 
fran each map were compared using a "difference of 
proportions" test to determine if there was a 
significant difference between them. There 
definately was a significant improvement in 
classification accuracy using thermal channel 6 data 
The mapping accuracies for each class were also 
compared and tested for significant differences at 
0.05 level. In these particular cases a "t" 
statistic was computed rather than a "z" value since 
the number of sample sites per class was lower than 
50. The results of these tests indicate that there 
is a significant improvement in detecting asphault, 
coal, water, and sparse vegetation by using thermal 
thresholds. The most dramatic improvement was 
observed in the water and inert classes. CLUS67 
maintained a high degree of classification error 
between water and coal. Water was also carrmonly 
classified as dark inert material and asphault. 
These descrepancies were almost entirely eliminated 
by use of the thermal data. 
Another significant advantage of using thermal 
thresholds was in discriminating between cool 
agricultural areas and the warmer urban and natural 
grass cover types. Even though a minor confusion of 
classes is still evident after using a combination of 
multispectral and thermal classification techniques, 
the land cover map of the Salt Lake study area 
derived from this study is a substantial improvement 
over previous TM, IMS, and MSS classifications of 
urban areas. 
REFERENCES 
Card, D.H. 1982. Using known map category marginal 
frequencies to improve estimates of thematic map 
accuracy. Photogranmetric Engineering and Remote 
Sensing 48:431-439. 
Carlson, T.N., & F.E. Boland 1978. Analysis of 
rural-urban canopy using a surface heat 
flux/temperature model. Journal of Applied 
Meteorology 17:998-1013. 
Chavez, P., Jr., S.C. Guptill, & J.A. Bowell 1984. 
Image processing techniques for Thematic Mapper 
data. ASP-ACSM Technical Papers Annual Meeting, p. 
728-743. 
Clark, J. 1980. Ihe Effect of Resolution in 
Simulated Satellite Imagery of Spectral 
Characteristics and Computer-Assisted Land Use 
Classification. NASA, Jet Propulsion Laboratory 
Report 715-22, Pasadena, CA. 
Quattrochi, D.A. 1983. Analysis of Landsat-4 
Thematic Mapper data for classification of the 
Mobile, Alabama metropolitan area. Proceedings 
17th International Symposium on Remote Sensing of 
Environment. Ann Arbor, MI, p. 1393-1402. 
Sheffield, C. 1985. Selecting band combinations 
from multispectral data. Photogranmetric 
Engineering and Remote Sensing 51:681-687. 
Todd, W.J. 1978. A Selective Bibliography: Remote 
Sensing Applications in Land Cover Inventory Tasks. 
Technicolor Graphics Services, Sioux Falls, SD. 
Wang. S.C. 1985. Ihe potential of Landsat Thematic 
Mapper data for applied research. University of 
New Orleans, New Orleans, LA. (Manuscript for 
article) . 
Ward, J.H., Jr. 1963. Hierarchical grouping to 
optimize an objective function. Journal of the 
American Statistical Association 58:236-244. 
Welch, R. 1982. Spatial resolution requirements for 
urban studies. International Journal of Remote 
Sensing. 3:139-146.