In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Yol. XXXVIII, Part 7B 
495 
where v is the n-dimensional vector for each pixel position (i,j) 
composed by the multispectral values of the n pan-sharpened 
bands, while V is the correspondent one for the reference 
bands. High values of SAM are related to high spectral 
distortion, but the radiometric distortion is not taken into 
account by the SAM index. A global evaluation of the image 
can be obtained by averaging the SAM values of all the pixels. 
Together with the ERGAS and SAM indexes, the usual RMSE, 
PSNR and Correlation score indexes have been also evaluated. 
The overall quality of a pan-sharpened image can be assessed 
by using a quality budget composed by a combination of more 
indexes, such as that proposed by Thomas and Wald (Thomas 
and Wald, 2007), but in the context of this paper the score 
indexes have been considered separately, in order to analyse 
the possible correspondence between the related ranking and 
the visual evaluation. 
5.2 Qualitative evaluation 
The qualitative evaluation has been performed on the full 
resolution pan-sharpened images by means of a visual 
inspection carried out by skilled photointerpreters, relying on a 
landslide inventory obtained by usual photointerpretation 
techniques on aereophotos of the study area. A suitable subset 
of the landslides present in the landslide inventory has been 
chosen, in order to consider the most significant ones. The 
analysis has been performed by considering the characteristics 
that allow the photointerpreter to map the landslide, such as 
quality of the linear features and textures, contrast and colour; 
the evaluation results have been averaged on the adopted subset 
and a related global quality evaluation has been finally 
produced. 
6. RESULTS 
The quantitative assessment of the GSA_CA, the GSG, the GS, 
the GIHS, and the PC fusion techniques have been performed 
by evaluating the ERGAS, the PSNR, the SAM, the RMSE and 
the Correlation Coefficient (CC) score indexes between the 
fused images and the reference ones. The adopted study area 
has not a rectangular shape, but the evaluation of the 
aforementioned quality indexes has been performed on a square 
subimage, in which the pixels outside the study area have been 
masked and set to the zero value. As a consequence, the high 
number of corresponding zero values between fused and 
reference image introduce a bias in the score indexes. In order 
to overcome this drawback and to obtain values that are easily 
comparable among the different fusion methods, the resulting 
score indexes have been normalized in percentage of the best 
achieved value. A value of 100 is therefore assigned to the 
better method among the tested ones, and consequently the 
higher is the number the worst is the performance for the 
ERGAS, the SAM, and the RMSE, whereas for the PSNR and 
the Correlation value the worst methods are characterized by 
the lower values. The normalized score indexes achieved by the 
tested fusion methods are listed in Table 2, together with the 
performance achieved by the “expanded” image, that is a 
simply resampled image obtained by interpolating the MS 
bands to be used as a reference to evaluate the improvement 
introduced by the pan-sharpening procedures. 
ERGAS 
% 
SAM 
% 
RMSE 
% 
PSNR 
% 
CORR 
% 
GSA 
-CA 
100.0000 
104.1346 
100.0000 
100.0000 
100.0000 
GSG 
104.0293 
115.8417 
195.1536 
98.6895 
99.9305 
GS 
112.2316 
121.4751 
114.3788 
96.9539 
99.7860 
IHS 
135.1306 
130.9472 
138.5633 
92.4371 
99.3656 
PC 
124.1912 
128.3047 
125.1895 
94.9071 
99.5846 
EXP 
134.9381 
100.000 
139.3647 
91.8082 
99.3706 
Table 2. Results of the quantitative evaluation. 
As expected, the GSA-CA and the GSG achieve the best 
performances for the most part of the adopted score indexes, 
even if GSG shows some problems in terms of RMSE, whereas 
the GS is found to be the best among the fusion methods 
implemented by ENVI. 
From the visual point of view, the results have been 
summarized in the table 3, in which for each studied fusion 
technique the resulting quality in terms of the characteristics 
used by the photointerpreter to map the landslide are assessed 
together with a global judgment of the fused image; the 
evaluation is provided by means of a detectability rating scale, 
based on a five levels ranking, namely: 5 (insufficient), 4 
(poor), 3 (medium), 2 (good), 1 (excellent). 
Features 
Texture 
Contrast 
Colour 
Overall 
GSA 
2 
3 
2 
2 
2 
GSG 
1 
1 
2 
2 
2 
GS 
1 
1 
2 
2 
1 
IHS 
3 
3 
2 
3 
3 
PC 
1 
1 
1 
2 
1 
Table 3. Qualitative evaluation 
By considering the scores of Table 3, the PC and the GS pan- 
sharpening techniques result as the best in the context of 
landslide detection among the tested ones. Also the GSG does 
not change linear features and texture, but is less satisfactory in 
term of colour quality, whereas GSA slightly suffers for some 
changes in linear features and texture useful in landslide 
detection. 
7. CONCLUSIONS 
Five pan-sharpening techniques, the Gram-Schmidt Global 
Adaptive (GSG), the Gram-Schmidt Adaptive - Context 
Adaptive GSA-CA, the Gram-Schmidt method (GS), the 
Principal Component method (PC) and the generalized-IHS 
(GIHS) method have been tested on a IKONOS multispectral 
data set acquiied over Umbria region in Italy, and the quality of 
the resulting pan-sharpened images have been compared 
quantitatively and qualitatively in the specific context of 
landslide detection task. From the quantitative point of view, 
the synthesis property introduced by Wald et al. (Wald et al.