Table 4. Bandwidth in microns and minimum and maximum 
radiance values (spectral radiance for TM) for 
calibration of Landsat data processed by EROS Data 
Center (Landsat 4 & 5, Esrange and Fucino also). 
BANDWIDTH Rmin Rmax 
Landsat 
1 
MSS 4 
0.5 - 0.6 
0 
2.48 
MSS 5 
0.6 - 0.7 
0 
2.00 
MSS 6 
0.7 - 0.8 
0 
1.76 
MSS 7 
0.8 - 1.1 
0 
4.00 
Landsat 
2 
(22 January 1975 - 16 
July 1975) 
MSS 4 
0.5 - 0.6 
0.10 
2.10 
MSS 5 
0.6 - 0.7 
0.07 
1.56 
MSS 6 
0.7 - 0.8 
0.07 
1.40 
MSS 7 
0.8 - 1.1 
0.14 
4.15 
Landsat 
2 
(after 16 July 
1975) 
MSS 4 
0.5 - 0.6 
0.08 
2.63 
MSS 5 
0.6 - 0.7 
0.06 
1.76 
MSS 6 
0.7 - 0.8 
0.06 
1.52 
MSS 7 
0.8 - 1.1 
0.11 
3.91 
Landsat 
2 
(5 March 1978 
- 3_1 May 
1978) 
MSS 4 
0.5 - 0.6 
0.04 
2.20 
MSS 5 
0.6 - 0.7 
0.03 
1.75 
MSS 6 
0.7 - 0.8 
0.03 
1.45 
MSS 7 
0.8 - 1.1 
0.03 
4.41 
Landsat 
3 
(after 31 May 
1978) 
MSS 4 
0.5 - 0.6 
0.04 
2.59 
MSS 5 
0.6 - 0.7 
0.03 
1.79 
MSS 6 
0.7 - 0.8 
0.03 
1.49 
MSS 7 
0.8 - 1.1 
0.03 
3.83 
Landsat 
4 
& 1 
MSS 4 
0.5 - 0.6 
0.04 
2.38 
MSS 5 
0.6 - 0.7 
0.04 
1.64 
MSS 6 
0.7 - 0.8 
0.05 
1.42 
MSS 7 
0.8 - 1.1 
0.12 
3.49 
Landsat 
4 
& 5 TM (expressed as s 
pectral rad: 
TM 1 
0.45 - 0.52 
-0.15 
15.21 
TM 2 
0.52 - 0.60 
-0.28 
29.68 
TM 3 
0.63 - 0.69 
-0.12 
20.43 
TM 4 
0.76 - 0.90 
-0.15 
10.62 
TM 5 
1.55 - 1.75 
-0.04 
2.72 
TM 7 
2.09 - 2.38 
-0.02 
1.44 
exist, e.g. nearest neighbour, bilinear interpolation 
and cubic convolution. The most important to achieve 
with resampling is that every pixel really fit into 
the new image or the map. Resampling by cubic convo 
lution yields geometrical precision that is superior 
to the other algorithms. When resampling a multi 
date set of images to a map it is important to do the 
things in correct order. If the map is considered to 
be poor in geometrical precision, it is important to 
carry out the resampling with a minimum of changes of 
the internal geometry of the image. The following 
procedure is to recommend: 
1) Resampling image to image of the multidate image 
set. In order to optimize the geometrical 
precision, a large number of ground control 
points, evenly distributed, and if necessary, in 
combination with a high order polynomial trans 
formation should be used. 
2) Resampling of the multidate set to a map projec 
tion. If the map is considered to have poor geo 
metrical accuracy, a low order polynomial trans 
formation should be used. The image is then 
adjusted to the control points, with a minimum of 
change of the internal geometry of the image. It 
is important to have the ground control points 
evenly distributed, and not to resample outside 
the points. If using a high order polynomial 
transform we will get a false impression of 
accuracy, since the residuals at each point may be 
rather low. If the map is accurate, a high order 
polynomial transform may be be applied. 
Resampling of thematic information e.g. classified 
images or digitized maps, can not be resampled by 
cubic convolution, since the algorithm involves an 
interpolation procedure, but require the nearest 
neighbour algorithm. 
The introduction of high resolution data recorded at 
considerable off-nadir view angles, like the SPOT, 
necessitates development of new procedures for 
geometrical correction than conventional resampling. 
A better satellite/sensor viewing geometry descrip 
tion will be available in the future and opens new 
possibilities for geometrical correction based on 
satellite orbital and sensor characteristics (Moccia 
& Vetrella 1986). 
4.3 Data reduction 
information 
work out. 
An alterr 
mation, whi 
classes anc 
(Schowengerc 
promising 
transformati 
vegetation n 
5. SATELLITI 
The digita] 
enormous amc 
be divided 
elements (He 
1) spectral 
variatioi 
image, 
2) textural 
distribu- 
measured 
3) context : 
ledge of 
The latt. 
denoted as 
between m 
automatic , 
spatial infi 
cation is . 
the locatio: 
is, on the 
textural, a 
means that 
point to be 
for interpr 
of context 
what machin 
Spectral 
efficiently 
to manual 
and speed. 
sensing ca 
complex da 
results ar 
work with c 
5.1 Multisp 
NOTE: Landsat MSS-bands are referred to as MSS 4, 5, 
6 and 7, instead of Band 1-4. 
means that data from some stations can not be 
regarded as useful for multitemporal analysis where 
radiometric calibration is necessary. 
A severe limitation to proper radiometric 
correction is atmospheric contamination of the image, 
e.g. the presence of haze. A common procedure for 
removal of the influence from haze is to assume that 
the longest wavelength band (e.g. MSS7) is unaffected 
by skylight or haze. A pixel that is supposed to 
represent zero DGL (complete shadow or clean water) 
is identified. Any offset of this pixel value is then 
assumed to represent atmospheric noise, and can be 
subtracted from all other bands. This procedure is, 
however, only valid under the two assumptions, that 
the atmospheric contribution is entirely additive 
and wavelength independent. 
4.2 Geometric correction 
All multidate analyses of imagery require accurate 
geometrical correction, carried out by resampling. 
The resampling is done either as image to image or 
image to map. Several algorithms for resampling 
The large quantities of data in modern remotely 
sensed images have highlighted the need for data 
reduction methods. One of the most commonly applied 
ones is the principal components analysis (PCA), or 
Karhunen-Loeve transformation. The aim of PCA in 
remote sensing is to remove redundancy in the multi- 
spectral image. It has been shown in many studies 
that the individual wavelength bands are highly 
correlated to each other. The reasons to this high 
correlation are (Schowengerdt 1983): 
1) natural spectral correlation 
2) topographic influence 
3) overlap in spectral sensitivity of sensors 
The first principal component will contain the 
largest possible amount of the total variance in the 
multidimensional data set. The second component will 
contain the largest possible amount of the remaining 
variance, and so on. Typical values of variance 
content of Landsat MSS data are 90 % in the first 
component and 5 % in the second one (Olsson 1985). 
This technique may be a valuable instrument in e.g. 
multispectral classification of rnultidate data sets. 
The dimesionality can often be reduced by 50 % per 
multispectral image. There is, however, a risk that 
useful information get lost in the transformation. 
PCA is scene dependent and general rules for how the 
Multispectr 
common remc 
tion algori 
been most f 
in common t 
unique. Mar 
Graetz et. 
assumption 
unsupervise 
classes we 
space, was 
Coiner (IS 
reliable me 
spectral c 
unnecessary 
informatioi 
vegetation 
Although 
weaknesses 
cases nece 
Several way 
1) Prior i 
expectec 
as pric 
ML-clasi 
ties a