Open Journal Systems

INTEGRAÇÃO DE VARIÁVEIS ESPECTRAIS E FORMA NA CLASSIFICAÇÃO DE IMAGENS DE ALTA RESOLUÇÃO UTILIZANDO REDES NEURAIS ARTIFICIAIS

Andréa Faria Andrade

Abstract



A classificação de imagens tem sido um importante processo para a
produção de mapas temáticos, e uma das principais fontes de dados para os
Sistemas de Informações Geográficas (SIG). Com o surgimento dos sensores de alta
resolução espacial, as aplicações não serão mais restritas aos estudos de clima e
monitoramento de florestas, mas sim cada vez mais nas áreas de SIG (redes,
telecomunicações, planejamento); mapas de arruamentos e cadastro; como também
no planejamento de traçados de rodovias, ferrovias e dutos. No entanto, para o
completo aproveitamento das novas imagens, com melhor resolução espacial,
abaixo de 5 metros, torna-se necessário disponibilizar de novas ferramentas, que se
estendam além da análise dos pixels isoladamente, possibilitando a modelação de
objetos, levando em consideração outros atributos da imagem, além dos espectrais.
Na classificação orientada ao objeto de imagens de alta resolução espacial, a
imagem é primeiramente segmentada e posteriormente os segmentos, que
correspondem a objetos, ou parte deles são classificados usando informações
espectrais, espaciais, como a forma ou a orientação dos segmentos, e a textura. O
presente trabalho se incumbe da inserção da forma no processo de classificação, por
considerar que ela é relevante na identificação de objetos presentes na imagem. Para
isto, o efeito de vários parâmetros de forma é avaliado usando uma imagem urbana
IKONOS, com resolução espacial de 1 metro. O algoritmo de classificação se
baseia no princípio das Redes Neurais Artificiais. A utilização de Redes Neurais
Artificiais emerge como uma alternativa vantajosa aos classificadores baseados em conceitos estatísticos, uma vez que nenhuma hipótese prévia sobre a distribuição
dos dados a serem classificados é exigida. A análise de forma comprovou ser uma
poderosa ferramenta para a discriminação de objetos que possuem a mesma
aparência espectral. Os resultados são comparados com métodos de classificação
convencionais baseados em dados puramente espectrais, para fins de avaliação do
ganho obtido incluindo a forma no processo.

Integration of spectral variables while classifying high resolution images by using
artificial neural networks

Abstract


Remote sensing image classification has been an important step for the
production of thematic maps, and one of the main data sources for the geographical
information systems (GIS). With the development of sensors of high spatial
resolution, the applications of remote sensing are no more restricted to climate
studies and environmeltal monitoring, but can be extended to other areas that need a
more detailed description of the earth's surface, such as telecommunication and
urban planning. However, for the complete use of the new images, with higher
spatial resolution, new tools are necessary, that enable to extend the analysis further
than the pixel based classification. Approaches like the object oriented image
analysis enable the modeling of objects, taking into account other attributes of the
image than the spectral ones. In the object oriented classification approach, the
image is segmented and later the segments, that correspond to objects, or parts of
them, are classified using spectral and spatial attributes, as the form or the
orientation of the segments, and the texture. The present work deals with the
insertion of shape information in the classification process, since it is considered
that it is important in the identification of the objects in the image. For this purpose,
the effect of the inclusion of several shape parameters into the classification schema
is evaluated, using an urban IKONOS image, with spatial resolution of 1 meter. The
classification algorithm is based on the neural network approach. The use of
artificial neural nets emerges as an advantageous alternative to the statistical
classifiers, since any previous hypothesis about the distribution of the data is not
demanded. The analysis of shape is a powerful tool for the discrimination of objects
that possess the same spectral appearance. The classification using spectral and
shape information produced a thematic image where the objects are better defined.
A comparison to other classification algorithms is also presented.