Open Journal Systems

APLICAÇÃO DE REDES NEURAIS NA CLASSIFICAÇÃO DE IMAGENS DE ALTA RESOLUÇÃO ESPACIAL E DADOS DO LASER SCANNER, USANDO UMA ABORDAGEM ORIENTADA A REGIÕES

Mosar Faria Botelho

Abstract



A classificação digital de imagens de sensoriamento remoto tem ganho,
nas últimas décadas, reconhecimento como técnica para suprir a escassez dos
mapeamentos temáticos. Atualmente, novos sensores, com maior resolução
espacial, encontram-se disponíveis no mercado, aumentando o potencial uso de
imagens de satélite. No entanto, estas novas imagens demandam novas técnicas de
análise, pois oferecem um maior grau de detalhe, o que possibilita a identificação de
objetos em função de suas propriedades espectrais e espaciais. Ao mesmo tempo,
novos dados espaciais vem sendo coletados por sensores aerotransportados, como
os dados do laser scanner. Estes dados altimétricos representam uma característica
particular dos objetos presentes na superfície da Terra, sua elevação, o que os torna
uma informação complementar valiosa na análise de imagens de sensoriamento
remoto.
O presente estudo aborda o tema da integração de dados derivados de um
levantamento laser scanner com imagens de satélite de alta resolução espacial,
Quickbird, para o mapeamento de áreas urbanas. Para isto, uma metodologia
orientada à análise de regiões na imagem é proposta. Inicialmente, o ganho obtido
com a inclusão dos dados do laser scanner é comparado com o ganho resultante do
uso de parâmetros espaciais derivados das imagens, através da análise da
separabilidade das classes e a classificação de áreas de treinamento. Os resultados
mostram que a contribuição da informação altimétrica é muito maior do que aquela atribuída aos descritores de forma. Em uma segunda fase, a utilização de redes
neurais artificiais como ferramenta para a integração dos dados espectrais e
espaciais foi avaliada, sendo que esta metodologia comprovou ser mais eficiente do
que outras abordagens tradicionais. O uso de redes neurais e os dados do laser
scanner aumentam a qualidade do mapa temático em regiões onde a informação
espectral não é suficiente para discriminar objetos diferentes.

Applying neural network by classifying images of high spatial resolution and laser
scanner data, using an approach to oriented region

Abstract


Remote sensing digital image classification has been recognized, in the
last decades, as a technique for thematic maping. Today, new sensors, with
enhanced spatial resolution, are available in the market, increasing the potential use
of satellite images. However, the new images demand new analysis techniques,
because they are richer in details, which should facilitate the identification of
objects, based on its spectral and spatial properties. At the same time, other spatial
data are being collected from airborne sensors, as laser scanner data. The altimetric
information represents a special characteristic of the objects present on the earth´s
surface, its elevation, which turns them a valuable information for the analysis of
remote sensing images.
The present study focuses on the integration of laser scanner data and remote
sensing images with high spatial resolution, Quickbird imagery, for urban mapping.
For this urpose, a region oriented aproach is proposed.Initially, the gain obtained
with the use of laser scanner is compared to the gain that results using of spatial
parameters in the classification process. For this urtose, the separability of the
classes was analysed and selected samles were classified. The experiments proved
that the contribution of laser scanner data is larger than that attributed to spatial
paremeters, form parameters. In a second step, the use of artificial neural nets as a
tool for the integration of spectral and spatial data was evaluated. This methodology
prooved to be more efficient than other traditional approaches. The use of neural
nets and laser scanner data increased the quality of the thematic map in areas where
spectral information is not enough to discriminate different objects.