SE- 
lity 
| to 
, it 
ea- 
see 
me 
the 
1er- 
ag- 
ion 
ing 
AR 
xel 
ler- 
has 
been calculated. Repeating the procedure for the 
entire interferogram, an irregular grid of 3D points 
(with the associated coherence) has been generated. 
The final geocoded coherence (30 m pixel size) has 
been obtained by resampling of the irregular grid. 
3.1.2 Amplitude Image 
The amplitude of the radar signal can be easily de- 
rived from the complex SAR images that contain, for 
each pixel, two components: 
I = A + sin D in-phase component 
Q = À + cos D in-quadrature component 
where: 
A signal amplitude, 
® signal phase. 
The amplitude is related to several geophysical pa- 
rameters of the earth's surface. The amplitude image 
employed for the classification has been derived from 
the complex image pair (from the master image) 
described in paragraph 2.1.1. In order to reduce the 
speckle, the original amplitude image (footprint of 
about 4 m [azimuth] by 20 m [ground range]) has 
been compressed in azimuth (incoherent average) 
generating a 4-look image. This image has been geo- 
coded (30 m pixel size) with the same procedure 
adopted for the coherence. 
3.1.3 Panchromatic SPOT Image 
The SPOT orthoimage employed for the classification 
has been generated at the Institute of Geodesy and 
Photogrammetry — Zurich Institute of Technology 
with the Helava DPW 770 (using the SPOT derived 
DEM described in paragraph 2.2). 
The original geocoded orthoimage, with a pixel size 
of 50 m, has been resampled to get the same grid (30 
m pixel size) of the amplitude and coherence images. 
3.2 Classification Procedure and Results 
The amplitude and coherence geocoded images have 
been processed in order to enhance their radiometric 
quality for the classification. First, they have been 
smoothed using a 3x3 low-pass filter (local filtering); 
then a 7x7 LUM (Lower Upper Middle) filter has 
been applied (edge preserving smoothing). 
Taking advantage of the DEM generated with the 
SPOT and InSAR data (see paragraph 2.3), an im- 
proved amplitude image has been derived considering 
the local incidence angle of the radar beam [Stussi 
and Beaudoin 1995]. Comparing the original and the 
radiometrically corrected images, the image quality 
improvement clearly appears. The improvement is 
particularly evident on hilly and mountainous terrain 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 
EST GE EIC UTR RR LTEM RETE EE QT Aen T t IGNI Hy a Ie ER 
OMM EE i PED NN A NP 
A. 8 £ 
where the relief-induced distortions (e.g. foreshort- 
ening) are reduced. 
The resampled SPOT orthoimage described in para- 
graph 3.1.3 has been directly employed for the classi- 
fication (without image enhancement). 
The RGB colour composite of the three images (red 
for SPOT, blue for the amplitude and green for the 
coherence) has been generated. It has been used for 
visual interpretation (see figure 3). 
The images have been introduced in a self-organising 
unsupervised classifier (ISOCLUST of the Idrisi 
software). The output clusters have been aggregated 
into the groups that fit the principal land use classes 
of the test area. 
The final output land use map (see figure 4) consists 
of four main clusters. The final- aggregated cluster 
have been compared with the reference land use map 
(Corine Land Cover) of the ORFEAS data set. In 
order to perform the comparison between the two 
maps, the number of classes of the reference map has 
been reduced to five (see figure 5): agricultural areas 
(arable land and permanent crops), forests (coniferous 
and mixed forests), water surfaces, fruit trees and 
urban and industrial areas. 
These are the comparison results: 
- 71 % of the pixels of the first cover class (agricul- 
ture) correspond to the first detected cluster. 
- 27 % of the pixels of the first cover class (agricul- 
ture) correspond to the second detected cluster. 
- 63 % of the pixels of the second cover class (forests) 
correspond to the second detected cluster. 
- 19 % of the pixels of the second cover class (forest) 
correspond to the first detected cluster. 
- 72 % of the pixels of the third cover class (water 
bodies) correspond to the third detected cluster. 
- 21 % of the pixels of the third cover class (water 
bodies) correspond to the second detected cluster. 
The last two classes (fruit trees and urban/industrial 
areas) do not correspond to any detected cluster. 
With the implemented classification procedure it is 
only possible to distinguish between 3 main cover 
types (agriculture, forest and water bodies) with a not 
very high accuracy. As mentioned previously, the 
employed data set is limited to very few layers. 
Significant improvements are expected extending the 
classification to more layers (in particular, using time- 
series of coherence images). 
4. Conclusions 
In this paper the fusion between optical and radar data 
for DEM generation and Land Use Classification is 
addressed. The data fusion between optical and inter- 
ferometric height data can improve both the accuracy 
(e.g. respect to the INSAR DEM in the rugged areas) 
and the completeness (e.g. respect to the SPOT DEM 
703