ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001 
ISPRS, Vol.3' 
■ 
I 
34 
Fig. 2 “ Analytical sequence ofper-field classification procedure” 
Ground Control Points (GCPs). A high (3rd) order of 
transformation for rectification (RMS value lower than 1 pixel) 
and a nearest neighbour resampling method on the image were 
executed, in order to consider the relief of certain sub-areas e 
the minor distortion of radiometric values in the row data 
respectively (Khan, Hayes and Cracknell, 1995). 
The spatial resolution of IKONOS image allowed individuation of 
nine land use classes (asphalt road, country road, arterial road, 
high density buildings, low density buildings, sown ground, 
uncultivated ground, Mediterranean bush, olive-grove). 
Generally, remotely sensed reflectance is related to land cover 
and not to land use, but in the present case each land use class 
was assumed to correspond to spectrally separable land covers. 
In the first step, the per-pixel classifier was trained on a 
representative sample of each of the land use classes by using a 
supervised maximum likelihood classification algorithm with 
equal prior probabilities for each class. 
This parametric classifier was selected because, by using the 
shape of the distribution of membership (represented by 
covariance) as well as the mean of the training data to identify 
each class, offered a very high general level of global accuracy. 
Then, the classifier modification procedure was performed, in 
order to improve the accuracy of the results with the correction 
of misclassifications by providing spatial context (digital 
topographic map). In this case the spatial context is the 
geometry of.a field that is an area in which only one land cover 
type is expected. 
All relevant fields boundaries was extracted from the digital 
cartography and processed as coverage in the GIS environment 
in order to perform the per-field classification in the image 
processing system (Fig. 2). Finally, a post-classification filtering 
with window width of 3x3 was executed to allow speckle 
reduction. 
Reference 
data set Classified Land Use 
1 
2 
3 
4 
5 
6 
7 
8 
9 
1 
3843 
0 
49 
0 
7 
27 
53 
16 
55 
2 
0 
159 
4 
12 
3 
1 
94 
350 
92 
3 
0 
37 
22437 
15 
55 
232 
751 
846 
1538 
4 
0 
0 
0 
5416 
0 
0 
301 
282 
424 
5 
0 
30 
27 
9 
1553 
631 
221 
124 
27 
6 
580 
0 
148 
0 
114 
5538 
1088 
518 
478 
7 
3 
9 
10 
239 
7 
4 
1717 
201 
150 
8 
0 
0 
3 
695 
0 
0 
81 
606 
254 
9 
0 
0 
7 
39 
0 
0 
41 
10 
824 
Overall Accuracy: 79.294% from 53085 observations 
Kappa statistic: 0.721 
Tab. 1 “Confusion Matrix of the Classification Map derived from IKONOS imagery of Fasano (Italy)” 
4. 2 Accuracy assessment 
Accuracy assessment determines the quality of the information 
derived from remote sensed data. To perform classification 
accuracy assessment correctly and to ensure objectivity and 
consistency, it is necessary to compare two source of 
information: (1) the remote sensing derived classification map 
and (2) the reference test information kept independent of 
training data. The relationship between these two sets of 
information is commonly summarised in a confusion matrix 
(Congalton, 1991). 
After the reference data set was collected from the randomly 
located sites, it was compared on a pixel-by-pixel basis with 
the present information in the classified satellite imagery. This 
source of information was utilised to validate the results of both 
classifications by calculating confusion matrices (Tab. 1), k 
statistics and overall accuracy. 
As the result of the whole procedure, inclusion of the 
topographic map information during the IKONOS image 
classification improved overall accuracy of the results (from 
68% to 79%). 
The per-pixe 
as a conseq 
to the utilisa 
Per-field da; 
some instan 
registration 
the selectior 
Flexibility < 
packages ; 
satellite im 
detail and 
images of 
researches 
problem to 
At present 
stored in i 
prove usel 
addition, re 
store and i