Use of features derives from class proportions in a pixel for the multispectral classification of remote sensing images / Utilização de atributos derivados de proporções de classes dentro de um elemento de resolução de imagem ("pixel")na classificação multiespectral de imagens de sensoriamento remoto

AUTOR(ES)
DATA DE PUBLICAÇÃO

1991

RESUMO

The spectral energy collected by the Remote Sensing instrumentation is the integration, called mixture, of the energy reflected or emmited by the objects, called primary components of the mixture, contained in a picture element. The land use classes in a scene can be described in terms of these components proportions, specially for forest targets. The objective of this issertation is to analyze the effect obtained in the automatic classification process when utilizing synthetic bands derived from components proportions in each pixel. A Linear Mixing Model is adopted and the estimated proportions are obtained by the use of the Constrained Least Squares and weighted Least Squares methods. The Landsat TM images used for the tests correspond to two reforested areas denominated "ITAPEVA" and "MOGI-GUAU". The analysis of the classification process is based on the Maximum Likelihood Algorithm using the gaussian hypothesis, and on methods of dimensionality reduction frequently employed in Remote Sensing. The obtained results show that, in general, by using a substitute attribute set (formed by the addition of the synthetic bands to the originals or only by the synthetic bands), a greater compression performance under the Principal Components and Canonical Analysis transformations is obtained. However, not significant improvement in the estimation of the average performance nor in the J-M Distance values between classes was obtained. In spite of this fact, the qualitative analysis of the thematic images provided important results: a) for the analyzed images it is possible to conclude that one should not use the substitute set composed by the addition of the synthetic bands to the originals; b) the best results are obtained by utilizing the synthetic bands only, provided that they are generated from the components which adequately represent classes in the scene and such that their proportions indicate the target structural differences. One observes, in this way, the importance of shade as a primary component for forest targets. The use of the synthetic bands only can be seen as a method for reducing the feature space comparable to the methods usually employed in Remote Sensing. The synthetic bands can also be useful for manual interpretation, due to their excellent visual effect produced by colour composites and also because their information represents physical concepts (proportions) easier assimilated than the classes spectral signatures.

ASSUNTO(S)

modelo linear de mistura pixels itapeva (sp)mogi-guacu mapeador tematico (landsat) maxima verossimilhanca reconhecimento de padroes satelites landsat land use spectral energy distribution targets classificacao automatica maximum likelihood estimates uso da terra

ACESSO AO ARTIGO

Documentos Relacionados