Sar image textural classification by autorregressive modelling / Classificacao textural de imagens SAR, por modelagem autorregressiva

AUTOR(ES)
DATA DE PUBLICAÇÃO

1998

RESUMO

The aim of this work is to present a method of textural feature extraction by autorregressive modelling to supervised SAR image classification. The classification is performed using a bidimensional processing. The autorregressive parameters are estimated using the Levinson-Durbin algorithm and they are used as the inverse filters elements. After that, two other approaches (whitening features by local autocorrelation function with lags (1,0) and (0,1) calculation and energy filter achievement) were applyed over the original filtered SAR image, as a way to improve the classification on the edges of the classes. The method was tested over JERS-1 (L band) SAR image and RADARSAT (C band) image from "Floresta Nacional do Tapajós" (FLONA). The identified classes for JERS-1 mosacic were Dense Primary Forest and Undulate Primary Forest. For continuous JERS-1 image and RADARSAT image the identified classes were Primary Forest, Secondary Forest and Bare Soil. For JERS-1 and RADARSAT images a Landsat TM image in color composition RGB-543, from the same area was used as ground truth. To experiment this modelling in other area a SAR-580 SAR image, from a Germany region, was also used, with identified classes Forest, Agricultural Area and Bare Soil. After the selection of samples from those classes (training and test regions) the method was tested, resulting a set of filtered bands from the original SAR image. Over this set of bands the Maximum Likelihood multibands classifier was used. The results were analysed using the confusion matrix.

ASSUNTO(S)

imagem de radar processos autorregressivos filtros de radar analise de imagens analise de regressão técnicas de imagens de radar processamento digital de imagens

ACESSO AO ARTIGO

Documentos Relacionados