- 1724 - 
Taiwan/Miller and Sung 
island of Taiwan at a scale of 1:25,000. Two Landsat frames from the same orbit path cover 
about 95% of the land area of the island. These frames have been designated as path 126 and 
rows 43 and 44, and only one clear set of imagery has been collected to date. This coverage 
occurred on 1 November 1972, and the data were utilized in this analysis. The site employed 
for this study is located on the western side of central Taiwan in the northernmost image and 
represents 2,100 square kilometers, or about 6% of the total island area. The study site repre- 
sents three of the 67 China Map Service 1:50,000 scale topographic maps of Taiwan. It begins 
on the west coast of the island at the Taiwan Strait and extends eastward through the coastal 
plains, terrace table lands, Taichung Basin and into the steep foothills of the central, north/ 
south-oriented mountain range. The area was selected because it contains a complete sam- 
pling of the types of land use practiced in Taiwan. Those portions of the study area contain- 
ing upland and mountainous terrain are generally forested, but some initial clearing is 
underway. Thus, the simultaneous comparison of the location of the agricultural land cover 
classes relative to forested cover and terrain roughness will provide a preliminary insight into 
the utility of the Landsat imagery for monitoring forest canopy conversion to agricultural 
lands. 
Approach and Results: The three 1:50,000 topographic maps of the test area were 
photographically enlarged to transparent overlays of precisely 1:25,000. Using an affine 
transformation, the digital imagery of these three areas was extracted, rectified and displayed 
to match the transparent map overlays (Fig. 1.1). Extensive experimentation was next under- 
taken to optimize both the classifications sought and the optimal spectral bands to be 
employed. The classification scheme was subsequently reduced to 17 cover types consisting 
of 2 urban, 5 agricultural, 3 forest, 3 barren land themes and 4 water depth/sediment classes. 
The training data for these tests and the subsequent classification runs were interpreted in 
Taiwan from airphotos. Since only one set of Landsat imagery was available for the island, 
the classification and spectral band optimization was restricted to 4 MSS bands and their 6 
ratios. The optimizations performed established that for the 17 land cover classes sought, a 
two-spectral space classification employing MSS bands 5 (Fig. 1.1a) and 7 (Fig. 1.1b) would 
be the most effective when using a maximum likelihood classification algorithm. The classi- 
fication maps which resulted were redisplayed as special theme maps to illustrate the spatial 
distribution of only the 5 agricultural cover types (Fig. 1.2a) and the 3 forest cover types 
(Fig. 1.2b). These classifications were verified by the air photointerpretation of the land 
cover for a 3 by 3 array of 9 Landsat cells taken at grid intersections of every 30 rows and 30 
columns in each map (Fig. 1.3). Close agreement can also be found in the comparison of the 
areal extent of agricultural and forested land cover in the classification maps and the esti- 
mates from the airphoto grid sample (Fig. 1.4). 
The upland or mountainous terrain can be readily observed from its shaded relief appearance 
on the easternmost (right-hand) map as rendered in the Landsat MSS band 7 graymap (Fig. 
1.1b). Comparison of this representation of upland physiography with the encroachment of 
agricultural land uses (Fig. 1.2a) into areas of forested land cover (Fig. 1.2b) illustrates the 
utility of Landsat for monitoring this process. The pending completion of a Landsat ground 
station in Japan and the possibility of a direct global satellite relay link between Landsat and 
NASA will provide the opportunity for at least yearly image coverage of Taiwan. Thus, a 
practical program for continual monitoring of forest watershed dehabilitation or reclamation 
appears feasible. 
Furthe 
Ref