International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
figure shows an (idealized) output of the fusion process, where 
the shape of the objects is identical with the panchromatic 
image, but the spectral behavior of the objects corresponds to 
the multispectral image. 
Fig. 1. Effect of Adaptive Image Fusion: (a) panchromatic 
image ; (b) multispectral image ; (c) fused 
multispectral image. 
Although AIF sharpens the multispectral image according to 
object edges found in the panchromatic image, this effect is 
limited to object edges that occur both in the panchromatic and 
the multispectral image. Edges appearing only in the panchro 
matic image will cause no significant effect in the multispectral 
band, while edges in the multispectral band that do not show up 
in the panchromatic image will get slightly blurred. Objects that 
are smaller than the original multispectral pixel size will only be 
sharpened, if their spectral reflectance is significantly different 
from their local environment. At the same time as object edges 
are sharpened, the area within an object will be smoothed. This 
effect can be seen as a pre-segmentation of the multispectral 
image. It reduces the occurrence of mixed pixels, thus 
improving the quality of a subsequent classification. 
3. APPLICATION 
3.1. Agricultural case study 
The AIF algorithm was first applied to a multisensor image 
acquired by the Indian remote sensing satellite IRS-1C. This 
system provides panchromatic images with a spatial resolution 
of 5.6m (PAN), and three multispectral bands (green, red and 
near infrared) with a resolution of 23.5m (LISS-2/3/4). The 
commercially available scenes are radiometrically preprocessed, 
panorama corrected and resampled to a pixel size of 5m and 
25m respectively. The acquired scene (path 30, row 35, quadrant 
A) was sensed on August 9, 1996 and covers an area of 
approximately 70x70 km 2 located in Upper Austria (centered on 
13°35’E / 47°58’N). The images were geocoded to the Austrian 
reference system (GauB-Kruger M31). The multispectral bands 
were resampled to 5m pixel size to match the resolution of the 
panchromatic image. 
Application of AIF requires the selection of two parameters, the 
normalised standard deviation and the size of the local window. 
The former can be estimated from the panchromatic image as 
described in section 2.1 The minimum size of the local window 
is the ratio between original multispectral and panchromatic 
pixel size. This minimum size offers the advantage of short 
computation time. However, test runs of AIF have shown that 
larger window sizes, such as 9x9 or 11x11, lead to significantly 
better results. This will increase the computation time of the 
program, but at the same time will reduce the number of 
iterations. This is due to the stronger effect of averaging, when 
larger window sizes are used. 
For the demonstration study a normalized standard deviation of 
0.01 and a window size of 11x11 were chosen. Each 
multispectral band was fused subsequently with the 
panchromatic image, thus leading to a synthetic multispectral 
image stack. Being the most commonly used merging technique, 
an IHS procedure was performed for comparison. 
Visual evaluation of the fusion results shows the differences 
between AIF and IHS merging (Fig. 2). The impression of the 
AIF image is that of a pre-segmented image, where the variation 
of grey levels within the objects is very low, while the objects 
are clearly separated from each other. Areas with a high 
variance in the panchromatic image, such as the village left of 
the center, appear like one homogenous object in the fused 
image. The IHS image has more textural information that 
supports visual interpretation, but also distorts the spectral 
information significantly, as can be seen in some of the 
agricultural fields. 
Multispectral classification of the fused images confirms the 
different effects (Fig. 3). A comparison of the classification of 
the original image and the AIF image, both classified with the 
same set of spectral signatures, shows the similar pattern of 
assigned classes. The delineation of objects is significantly 
better in the AIF image, and it contains less single pixel objects. 
Classification of the IHS image was performed with a different 
set of spectral signatures, as the spectral distortions did not 
allow using the signatures from the original multispectral image. 
The appearance of the classification is noisier, and in some areas 
differs significantly from the original classification. 
Quality assessment. The assessment focuses on how much the 
radiometry of the multispectral images is distorted by the fusion 
procedure. It is based on the idea that a synthetic image once 
degraded to its original resolution should be as identical as 
possible to the original image. This property is estimated by 
comparing the mean values and standard deviations of and the 
correlation coefficients between the degraded and the original 
images (Wald et al., 1997). In order to produce the degraded 
images, the fusion results were averaged applying a 5x5 filter 
kernel and resampled to 25m pixel size. Table 1 presents the 
results of the comparison between the degraded images and the 
original data. The first column shows the global mean (p) and 
standard deviation (a) of the original multispectral bands. The 
columns entitled AIF and IHS give the respective values of the 
merged products. 
The results of AIF show no significant differences in the mean 
values, but the standard deviation is slightly reduced. This is 
due to the averaging performed during the merging process. IHS 
merged images show a higher deviation in the mean values and 
an increase of the standard deviation resulting from the 
inclusion of textural information from the panchromatic image. 
It is interesting to note, that for channel 4 (NIR) the standard 
deviation has decreased during the IHS merge.