3b. Beijing 2008 
H„ 2005. Seed 
i potential use in 
national Journal 
t, Voi. 13, No.l, 
ACCURACY OF LACUNARITY ALGORITHMS IN TEXTURE CLASSIFICATION OF 
HIGH SPATIAL RESOLUTION IMAGES FROM URBAN AREAS 
Algorithm for 
:. Letters 9, pp. 
essing, analysis, 
for recognition 
M. N. Barros Filho 3 ’ *, F. J. A. Sobreira b 
a Faculty ESUDA, R. Barao de Itamaraca, 460/1802, 52020-070, Recife, Brazil 
- mbarrosfilho@gmail.com 
b UNICEUB - Centro Universitario de Brasilia, Brasilia - DF, Brazil 
- fabiano.sobreira@gmail.com 
Commission III, WG III/4 
KEY WORDS: Accuracy, Algorithms, Texture, Classification, High Resolution, Urban 
ABSTRACT: 
Lacunarity based measures can be described as texture recognition approaches that provide a flexible yet theoretically consistent 
mean of characterizing the morphology of urban spatial patterns across different scales. This paper proposes a comparison between 
the Gliding-Box and Differential Box-Counting algorithms based on the concept of lacunarity to recognize and classify textures of 
urban areas with different inhabitability conditions through the analysis of binary and grayscale images from 30 ® Quickbird sensor 
image samples from Recife (Brazil), captured in October 2001, with 250 x 250 meters in size. Results show that the Differential Box 
Counting algorithm applied in grayscale images improves the discrimination between textures from urban areas with different 
inhabitability conditions, and it reveals a strong correlation between urban morphology and socioeconomic patterns. 
1. INTRODUCTION 
Satellite images are rich sources of information about earth 
surface, and provide wide coverage, frequent updates and 
relatively low costs. Over the last years, with the significant 
improvements in image resolutions and digital image 
processing techniques, remote sensing activities have been 
increasingly focused on urban environments. The development 
of techniques to extract and classify high resolution satellite 
images is an important task in urban planning, especially those 
from developing countries which lack reliable and accurate 
maps to represent their rapid and informal growth. 
Despite current advances in remote sensing technologies, 
accurately classifying high spatial resolution satellite images 
into urban land-use classes remains a challenge. Urban areas are 
composed of objects with different forms and materials, and 
high spatial resolution images from these areas lead to a more 
complex combination of surfaces with different spectral 
reflectance (Myint, 2007). 
Besides that, high spatial resolution images have, in general, 
low spectral resolution (a small number of bands), and this may 
make difficult the distinction between different urban features 
with similar reflectance in the visible wavelength (Donnay et al, 
2001). Traditional per pixel spectral classification algorithms 
like maximum likelihood do not take into account the spatial 
arrangement of neighborhood pixels, and spectral attributes 
alone cannot provide good classification results. 
Texture is a description of the spatial variability of pixel tones 
in a digital image, and it may improve image classification of 
urban areas. Texture analysis of digital images aims to 
recognize and to distinguish spatial arrangements of gray levels 
values, based on methods which measure the spatial variability 
of pixel tones in an image. The higher the variability, the less 
homogeneous or uniform will be the image texture (Barros 
Filho and Sobreira, 2005). 
A texture pattern is scale dependent. It may varies significantly 
according to the size and spatial resolution of a digital image. A 
very small image may contains parts of a pattern, and it may not 
be able to characterize the whole pattern, whereas a large image 
may be composed of more than one single pattern and could not 
be able to properly describe it as well. In the same way, a pixel 
in a low spatial resolution image may represent an integrated 
sign of many patterns smaller than the pixel size. As the spatial 
resolution increases the image pixels could become smaller than 
the analyzed pattern, generating spectral noises that degraded 
image classification (Mesev, 2003). 
Lacunarity based measures provide a flexible yet theoretically 
consistent mean of characterizing texture patterns across 
different scales. Experiments with binary ® Quickbird images 
from the city of Recife (Brazil) showed that it is possible to 
distinguish texture patterns of urban areas with different 
inhabitability conditions. Urban areas with better inhabitability 
conditions had high lacunarity values than those with worse 
conditions. These differences, however, tend to decrease as the 
scales become finer (Barros Filho, 2006; Barros Filho and 
Sobreira, 2007). These results are coherent with another 
experiment done in the city of Campinas (Brazil), when textures 
of binary ® Ikonos images were analyzed (Barros Filho and 
Sobreira, 2005). 
Experiments with binary ® CCD/CBERS-2 image data of 20 m 
spatial resolution from the same city (Recife, Brazil), showed 
an opposite relation between lacunarity and inhabitability: in 
general, image from urban areas with better inhabitability 
conditions have lower lacunarity values than those from poor 
Corresponding author. 
417