15
ie Sozialwissen-
C.H. Hull. Eine
ersionen 8 und
gart - New York,
Satellitenbild
aßstab
lume "Hohe
des österrei-
gramms Hohe
1 map enclosed,
ag Wagner.
9. Geologie des
il. Abhandlun-
r Reichsstelle
25, Issue 1,
ote sensing
ARSS 1982,
Anwendungen
rrain Model
österr.
Zeitschrift
ation - The
manual digi-
RTO IV, Nachr.
n, Ser . II,
!ain (IFAG ) .
on im Einzugsge-
estlich des
: Untersuchungen
en Tauern 1974 -
rhaushalt. Ver-
MaB-Hochgebirgs-
me 3, P- 35 - 67,
Universitätsver-
g und Erprobung
rischen Auswer-
ktralen Zellen-
g/digital-ge-
oma Thesis,
raz, 228 p.
gen zu den Bo-
und Umgebung
.000. In: MaB-
eröffentli-
hgebirgspro-
7, p. 23 - 28,
Universitäts-
ilogische Kenn-
Serien längs
biet des
r (Hohe Tauern,
Öffentlichungen
ogramms "H . T." .
g und Systema-
ption. Hercynia
derkarten
bung" 1 : 5.000
" 1 : 5.000 des
gramms Hohe
"Hohe Tauern",
erreich MaB-
ume 7, p. 29 -
uck, Universi-
Hütter 1986.
formation system.
Dibag Report,
age Processing
z Research Center.
Symposium on Remote Sensing for Resources Development and Environmental Management / Enschede / August 1986
Digital processing of airborne MSS data
for forest cover types classification
Kuo-mu Chiao, Yeong-kuan Chen & Hann-chin Shieh
National Taiwan University, Taipei
ABSTRACT: The purpose of this study is to find optimal band combinations of airborne MSS data for forest cover
types classification. An eleven-channel airborne multispectral scanner was used to collect data. Processing of
the MSS data was achieved through the use of IDIMS. Forest cover types were interpreted with a hybrid super
vised/unsupervised approach. Eight original bands of airborne MSS data and enhanced data such as principal
component transformations, ratio images, spatial filtering data, resampling data and mixed bands were subjected
to standard clustering and classification techniques. Among the numerous band combinations of MSS data, forty-
seven best band combinations with highest values of average divergence and minimum divergence were selected,
and classified on IDIMS by the maximum likelihood classifier. The optimal band combination for forest cover
types classification provided the most accurate and detailed classification results while minimizing computer
time and man-hours. Findings of this study are summarized as follows: (1) the best band combination is the
combination which contains more bands; (2) resampling and spatial filtering techniques increase 7-10% classifi
cation accuracy; and (3) Because of the inherent property, the principal component transformations, ratio
images and mixed bands are insufficient to improve the classification accuracy.
1 INTRODUCTION
Resource surveys are carried out through the use of
an airborne multispectral scanner rather than through
the satellite data. The airborne multispectral
scanner, which possesses the capabilities of fine
resolution, narrow wavelength band and flexible
scanning time, is more adequate on land cover types
mapping than the Landsat system in an area with many
vegetation types to be classified such as in Taiwan.
The goal of this study is to find the optimal wave
length band combination on land cover type classifi
cation in forested area. Experiences have indicated
that when working with many wavelength bands, maximum
accuracy in classification can be obtained by using
all wavelength bands available. However, this requires
a marked increase in computer time. It is therefore
frequently desirable to reduce computer time by
utilizing only limited wavelength band in the classi
fication procedure. The problem is which combination
of wavenlength bands would be the optimum set to use
in the classification.
2 MATERIALS AND METHODS
2.1 Data utilized
The airborne multispectral scanner data used in this
study were collected by a DS-1260 airborne MSS system
with 11 channels in the Chi-tou tract of the Experi
mental Forest of National Taiwan University in central
Taiwan on November 21, 1982. A computer compatible
tape (CCT) containing MSS data in band 4 to band 11
was used in data processing (table 1).
In order to produce images adequate for forest
cover types classification, image enhancement tech
niques were used to create image products which
protray spectral pattern representing a variety of
surface features and cover types. The enhanced data
are principal component transformations, ratio images,
spatial filtering data and resampling data.
2.2 Processing techniques
The image processing techniques were implemented
Table 1. Airborne MSS spectral bands for data
processing
Channel
Wavelength range (pm)
4
0.50-0.55
5
0.55-0.60
6
0.60-0.65
7
0.65-0.69
8
0.70-0.79
9
0.80-0.89
10
0.92-1.10
11
8.50-13.0
through the use of the Interactive Digital Image
Manipulation System (IDIMS) at National Central
University, Chungli, Taiwan, Republic of China.
As a first step in processing the imagery, the
MAGNIFY, SCANFIX and REGISTER functions in IDIMS were
used to geometrically correct the airborne MSS data.
Transformations were completed to correct the syste
matic and non-systematic errors such as aspect ratio
error, tangential scale distortion, altitude varia
tion and attitude variations. The non-systematic
distortions are not predictable, 18 widely scattered
ground control points were selected to determine the
geometric transformations required to correct the
image.
To generate the enhanced imageries, the principal
component transformation was first performed on the
8 original bands data. In accomplishing this proce
dure, function KLTRANS in IDIMS was applied. The
principal component 1 (PCi) contributes 82.77% of
the total variance by its eigenvalue; PC2 contributes
12.24%; and PC3 contributes 4.61%. Together, PCi,
PC2, and PC3 already account for 99.62% of the total
variance of the 8 bands data.
The second enhanced data is ratio images. By the
function ADD, DIVIDE, POWER, SCALE, HIST0G and
CONVERT in IDIMS, the 8 original airborne MSS .bands
can be combined to produce several dozens band ratios.
Among them, 8 ratios were selected for data processing
experientially. They are MSS10/MSS8, MSS10/MSS6,
MSS7/MSS9, MSS9-MSS6/MSS9+MSS6, MSS10/MSS5, MSS9/MSS11,
MSS4/MSS7, and MSS5/MSS11. In these band ratios, some
are good for vegetation classification; some eliminate
