> which 
vithin 
le in the 
is matrix 
nd or 
*. Only 
ng the 
y be as- 
omplex 
r of the 
he over- 
nges in 
em to- 
1dscape 
ils, etc.). 
J; 
tes which 
e images 
hese 
e-level 
sent both 
it agri- 
by that 
urface 
ndscape 
rlays of 
itribu- 
ie Land- 
(Table 
resenta- 
posite 
sisting of 
| assess 
propor- 
e area 
72, 1975 
—- 1755 - 
Thailand/Miller, Nualchawee and Tom 
The overall performance of the various classifiers and training set approaches tested were 
directly evaluated on the 27 January 1973 Landsat image. A complete classification map of 
the 29,290 cells common to all data sets was prepared for each case and checked cell by cell 
against the 1972 airphoto-derived forest cover map*. The accuracies achieved (Table 7.1 and 
Fig. 7.8) do not appear very high but this will always be the case when dealing with map veri- 
fication accuracy rather than the commonly employed training set accuracy. The latter 
verification scheme is much more subjective or even operator manipulated. The verified map 
classifications achieved the best overall accuracy of 55.4% with grid sampled training sets and 
proportional or apriori training sets. It should be noted that the percentage of correctly 
mapped points in such a complete map verification scheme cannot be expected to approach 
100%, as the airphoto interpretation maps of land cover have their own inherent accuracy 
which is less than 100%. For example, assume that the land cover map derived from the 
airphotos is 70% correct; that is, if 10 random cells were checked on the ground, 7 of the 10 
would have been assigned by photointerpretation to the correct cover type. This is a reason- 
able assumption in such a complex tropical forest canopy. The Landsat image classification 
map may also be 70% correct relative to ground conditions. Thus, verification of the Landsat 
classification map against the airphoto map will produce a 70% times 70% verification accur- 
acy, or 49%. This value is very similar to the accuracies achieved in this study, as noted above 
and in Table 7.1. 
Further Information: 
References 
Miller, L. D., K. Nualchawee and C. Tom. 1978. Analysis of the dynamics of shifting 
cultivation in the tropical forests of northern Thailand using landscape modeling 
and classification of LANDSAT imagery. Proc. Twelfth International Symposium 
on Remote Sensing of Environment (also NASA/Goddard Space Flight Center, 
Tech. Memo. 79545, Greenbelt, Maryland). 19 p. 
Wacharakitti, S., L. D. Miller and C. Tom. 1975. Tropical forest land use evolution/ 
twenty year landscape model with inputs from existing maps, historical airphotos 
and ERTS satellite imagery. Colorado State University, Environmental Engineering 
Technical Report 1, Ft. Collins, Colorado. 217 p. 
Nualchawee, K., L. D. Miller and C. Tom. 1978. Spatial land-use inventory, model- 
ing and projection in northern Thailand with inputs from existing maps, airphotos 
and LANDSAT imagery. Texas A&M University, Remote Sensing Center, Techni- 
cal Report, College Station, Texas. 220 p. 
Experimenters 
Lee D. Miller, Texas A&M University, Remote Sensing Center, College Station, Texas 
77840 U.S.A. 
Kaew Nualchawee, Colorado State University, Department of Civil Engineering, Ft. 
Collins, Colorado 70523 U.S.A. 
Craig Tom, HRB-Singer, Inc., Environmental Analysis Group, Science Park, State 
College, Pennsylvania 16801 U.S.A. 
*Note that December 1972 airphotos were used in preparing the 1972 forest cover map and are separated 
by about one month from the Landsat scene classified.