Working concept
[ I 1
Illumination effect | [Additional channels | | Datafusion |
Topographic Ratio-Channels Brovey-
Normalisation Transformation
Illumination- Difference-Chan. Sigma-
masking Adaptive Filter
Textur LMM-Verf.
ied
Classification approaches
|
Verification of classification
|
Landscape Indices calculated from
the best classification
wii
Illumination Masking
Separate classifications of illumination masks were
carried out. The illumination values were calculated using
the digital elevation model and the sun parameters during
the data take. Than three respectively two illumination
masks were separated for the classification.
a) 3 Masks
(0 -60%, 61-80%, 81-100%) Illumination
b) 2 Masks
(0 -60%, 61-100% Illumination)
Every mask was classified separately, combined to the
classification image and analysed in the accuracy
assessment process.
Additional Channels
In addition to the illumination masks ratio-, difference and
texture channels were computed from the original data
set to enhance the classification process. The
enhancement was judged by the mean separability of the
training data set using the Euclidean distance.
Result of this investigation was the selection of the
difference channel (xs3-xs2) and (TM4 - TM3)
respectively. From the ratio channels tested here only
one ratio was used in the layerstack for SPOT-XS
data(NDVI). In the layerstack of LANDSAT TM there were
additionally two other channels integrated, the IR-Index
and the greenness channel of the Tasseled Cap
Transformation. All this channels produced a better
separability of the training data set than the original input
channels.
Texture Channels
For the texture analysis six different texture measures
were tested ( Homogeneity, Contrast, Dissimilarity, Mean,
Standarddeviation and Angular Second Moment ). The
input parameters for the texture measures were tested for
3x3 window 5x5 window 7x7 window
and four neighbourhoods ( 0,1; 0,2; 1,0; 2,0).
The computation of texture measures was done with PCI-
image processing software. The cooccurrence matrix
showed improvements in the discrimination for the
Angular Second Moment listed below for the SPOT-XS
data set from 1992 and 1995.
ASM - N1,0 - Ch2 - W3x3
ASM - NO,2 - Ch2 - W3x3
ASM - N2,0 - Ch2 - W3x3
ASM - N02 - Ch1 - W3x3
ASM Angular Second Moment
N Neighbourhood
Ch Input-Channel
Ww calculation window for cooccurrence
In contrast to the SPOT-XS texture measures the
LANDSAT TM texture parameters and measures for the
mean separability were different. In the list below you find
the four best parameter combinations.
HG - N2,0 - Ch3 - W3x3
HG - N2,0 - Ch6 - W3x3
HG - N02 - Ch3 - W3x3
HG - N2,0 - Ch3 - W5x5
HG Homogeneity
N Neighbourhood
Ch Inputchannel
W calculation window for cooccurrence
The additional channels ( difference-, ratio-, texture
channels ) were concluded in a layerstack with ten and
eleven channels for SPOT and LANDSAT respectively.
The number of pixels in the training areas was around 30
to 100 pixels. Based on the restriction of available pixels
three channel combinations from the layerstack were
classified during the classification process.
Data Fusion
In the recent literature different procedures for fusion of
multispectral and panchromatic satellite data are
available.
According to the papers from DE BETHUNE (1997) the
procedures of data fusion will change the radiometric
values of the original image more or less strongly.
While the Brovey-transformation will result in a clearly
change of the radiometric values of the input image, the
372 International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998
PC
Thi
nat
cor
Ad
Im:
apr
Ad.
USi
bor
the
insi
sto
The
Inte
Syr
Loc
The
of n
(ij)
(iJ
low
the
divi
resi
fixe
The
trait
