International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
113 
Fig. 4. Fusion of ATM image data, (a) original image (6m), 
© ÔBIG 1991. (b) degraded image (18m). (c) AIF 
imaee f6ml. 
geocoded with respect to a digital terrain model in the Austrian 
reference system, thus no geometric distortions appear between 
the images. 
The derivation of the forest mask is based on thresholding of the 
single Landsat TM image bands. Applying this technique on the 
original image bands leads to a reliable forest mask, however the 
geometric resolution of this mask is limited by the pixel size of 
30m. On the other hand, the panchromatic image that would 
provide the better spatial resolution does not offer the spectral 
information needed to separate the forest areas from other land 
cover objects. The fusion of both data sets was expected to 
result in an improved forest mask. 
The AIF was applied to the two data sets using a window size of 
15 x 15, 7 iterations and a variance decreasing iteratively from 
0.025 to 0.01. Next the thresholding technique was applied to 
the resulting multispectral fusion product using the same 
thresholds as for the original Landsat TM bands. As expected, 
the resulting forest mask delineates the forest areas significantly 
better. 
Fig. 5 shows the two vectorised forest masks superimposed to an 
aerial colour infrared photograph of the area. Visual evaluation 
of the results shows the improvement gained from the fusion 
approach. The borderlines of the forest mask are smoother and 
follow the actual border of the forest areas. Due to the pre 
segmentation effect of the AIF, only areas of a certain minimum 
crown coverage are recognized as forest. 
4. SUMMARY AND CONCLUSIONS 
The case studies demonstrated the benefits and limitations of 
the AIF algorithm. The AIF was designed as a pre-processing 
tool for subsequent numerical classification. Its advantage lies 
in the pre-segmentation of the fused image, resulting in a low 
variance within the objects and sharp borders between the 
objects. The inclusion of highly textured areas from the 
panchromatic image is not supported by the technique, but the 
spectral characteristics of the multispectral image are preserved 
to a high extent. 
The first case study demonstrated the use of AIF for agricultural 
applications. Objects of interest are the single fields that 
represent the different agricultural crops of the area. These 
objects are expected to be of such a size that their spectral 
characteristics are represented even by the lower resolution of 
the multispectral image. Application of AIF improves a 
subsequent classification by sharpening the borders of the 
fields, thus reducing the number of mixed pixels and allowing 
for a more precise estimation of the actual areas covered by the 
single crop types. 
The urban case study proved the applicability of AIF to very 
high spatial resolution multispectral data. By degrading the 
multispectral image to a lower spatial resolution and fusing the 
result with an artificial panchromatic image, a comparison