fast and suitable system to detect and accurately measure the 
artificialized land of the great mega-cities through the use of 
low-resolution satellite images (Landsat 7). 
1.3 Specific objectives 
The specific objectives can be summarized as follows: 
The first aim of this paper is to delimitate the study 
area of the selected cities. A “window” of 45,000 
square kilometres of each megacity will be studied in 
a detailed level. 
* The second aim of this paper is the classification and 
interpretation of archived satellite images for the 
identification of land covers. Rural and artificialized 
land (including roads and urban green) will be 
classified. Urban texture analysis plays an important 
role in image segmentation and object recognition for 
improving objects extraction and classification. 
* Another goal of the paper is to delimitate, measure 
and understand the urban continuum inside and 
outside the administrative boundaries. The 
morphological analysis of conurbation will serve to 
identify the core city from the surrounding 
countryside and to compare the different structures of 
the studied megacities. 
* Comparison between the different models of 
urbanization. This stage is directed towards analyzing 
from a morphological perspective the process of 
spatial occupation of the megacities. Landscape 
indicators will be used to define the pattern of 
urbanization of each megacity. Urban sprawl, 
monocentrism vs. polycentrism, fragmentation and 
others aspects of the urban structure will be studied to 
compare the selected megacities. 
Finally, our aim in this paper is to generate an indicator to 
present geometric properties and intrinsic morphology of urban 
settlements based on their characteristics and fundamental 
forms, and to develop a strategic urban model that guides 
sustainable development of selective “Megacities”, using 
innovative technologies such as Remote Sensing, Geographic 
Information Systems (GIS) and WEBGIS. 
2. METHODOLOGY 
2.1 Analysed data 
Data sources are based on satellite images from Landsat 7 
ETM+ and the GLS (Global Land Survey) 2000 Collection 
brought together, which provides multispectral images of 30 
meters, plus a 15 m panchromatic image, in Geotif format. 
2.2 Land Use Classification Efficiency According to Per- 
Pixel Based Approaches 
The complexity of image classification techniques can range 
from the use of a simple threshold value for a simple spectral 
band to complex statistically based decision rules that operate 
on multivariate data. Classification approaches can be 
implemented to classify the total scene content into a limited 
number of major classes or can also be implemented to 
distinguish one or more specific classes of terrain (such as water 
bodies, artificialized land or other types). Pixel-based 
approaches have been developed and are widely used in remote 
sensing image processing and classification. Supervised 
    
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
   
classifications, such as Maximum Likelihood, Minimum 
Distance, Mahalanobis Distance, and Parallelepiped 
Classification, have been used. 
  
Figure 1.Per pixel classification approaches 
A number of issues have to be taken into consideration when 
selecting a suitable classifier: 
* Different training strategies may produce different 
classifications results. The training size, the image 
resolution, and the degree or autocorrelation inherited 
in each class influenced the performance of different 
training strategies. 
° The size of the training set is important in influencing 
supervised classification results when the single-pixel 
training strategy is applied. The number of pixels 
required to extract training statistics vary for different 
classes with different spatial structures. For spectrally 
homogeneous classes, a small number of training 
pixels may be sufficient. But for spatially 
heterogeneous classes, a relative large number of 
pixels are likely to be required in order to extract 
representative statistics. 
*  Single-pixel may be implemented to avoid auto- 
correlation effects, but not always lead to more 
accurate classification results than other training 
approaches involving contiguous pixel selection. For 
spectrally homogeneous classes, the single-pixel 
training approach may be preferred. But for spatially 
heterogeneous classes, small-block training has the 
advantage of easily capturing spectral and spatial 
information and saves the analyst interaction time. 
* Using an overlay tools as an additional process 
increases the classification accuracy for all classes and 
tends to reduce the differences of classification results 
caused by training strategies at all levels when the 
pixel-base classifier is used. 
Among the most frequently used classification algorithms, the 
maximum likelihood method is generally preferred (Campbell 
1998, Avery & Berlin, 1992). It becomes the most commonly 
used classifier due to its higher accuracy levels. It is generally 
accepted that this is the more accurate form of classification if 
compare to parallelepiped and minimum distance algorithms 
(Curran 1985). In our experience, the overlay of maximum 
likelihood and minimum distance classifications allow good 
performance in per pixel-analysis.