Desenvolvimento de classificadores de máxima verossimilhança e ICM para imagens SAR / Development of maximum likelihood and ICM classifiers for SAR images

AUTOR(ES)
DATA DE PUBLICAÇÃO

1996

RESUMO

The purpose of this study is to implement, test and apply a Maximurn Likelihood Classifier (MLC) and a userfriendly contextual Markovian classifier, which use the distributions most appropriate to the Synthethic Aperture Radar (SAR) data. This study presents the main distributions to the SAR data and how several of these distributions arise from the multiplicative model. In order to achieve the proposed aim and to allow future applications of the developed methodology, the implementations have been done in a integrated system for SAR data processing, classification and analysis, with a structure which allows the addition of other models and techniques. The developed system is based on the statistical properties of the data and, besides the necessary functions for the classification, it also allows the determination of the basic statistics of the classes radiometries, the Chi-Square goodness-of-fit test in order to choose the most suitable distributions for these radiometries, the classification itself, and the evaluation of the results with the Kappa coefficient of agreement for the error matrices. The implemented contextual classifier is the Iterated Conditional Modes (ICM) algorithm. Its formulation for the parameter estimation was done for 8- and 12-coordinate neighborhoods. Tests have been done for the discrimination of three classes in SAR-580 and JERS-l images with different numbers of looks. The analysis of the results indicates that a more precise classification is achieved by using the distributions which are the most suitable for the classes, when compared to those obtained through the classic method that uses Gaussian distributions. Another conclusion is that the ICM algorithm presents results which are always superior to those obtained through the MLC classification, whichever distributions these radiometries may have. The classifications obtained from the ICM have Kappa values that are usually twice than those obtained with the MLC method. Therefore, when using the ICM algorithm, it is possible to achieve good results in the discrimination of classes such as primary forest, regeneration, and deforestation in rainforest areas, as it is the case of the JERS-l images used in this study. The developed system has also some operations which assist the classification of images in general (color table modification and edition, histogram equalization, sample manipulation, decorrelation of observations, arithmetical operations, etc.), and other operations specific for SAR images (speckle filtering, equivalent number of looks estimation, selection of image type and modelling, etc.).

ASSUNTO(S)

radar de abertura sintética (sar) algoritmos deforestation synthetic aperture radar erros forest imagem de radar maxver algorithms icm maximum likelihood estimates florestas processamento digital de imagens modas condicionais iterativas desmatamento digital image processing máxima verossimilhança

ACESSO AO ARTIGO

Documentos Relacionados