Segmentação multi-níveis e multi-modelos para imagens de radar e ópticas / Multi-levels and multi-models segmentations for radar and optical images

AUTOR(ES)
DATA DE PUBLICAÇÃO

2005

RESUMO

This research presents an innovative automatic segmentation method (SegSAR) designed for radar and optical satellite imagery. The design of SegSAR is innovative because it encompasses (i) several common segmentation techniques such as region growing and merging, edge detection, and minimum area and homogeneity test, which are integrated in a pyramidal compression structure; (ii) a hierarchical segmentation procedure; (iii) a multi-level segmentation procedure; (iv) a multi-model homogeneity test; (v) a single band or multi-bands of satellite imagery option; and (vi) the generation of intermediate results at each level of the segmentation process. The processing of radar imagery is performed in intensity values; however, options of input image are amplitude, intensity, and dB. SegSAR is designed to represent imagery in two different models, (i) the cartoon model that is suitable for homogeneous regions, i.e., constant backscattering, of the image; and (ii) the texture model that is suitable for heterogeneous regions, i.e., variable backscattering. For segmentation of radar imagery, the cartoon model assumes that the data for each region has a Gamma distribution, and uses the coefficient of variation to determine the homogeneity degree of a region. Gamma distributed samples are used to obtain the critical coefficients of variation, at different levels of significance, through Monte Carlo simulation. Differently, the texture model uses only a comparison between a coefficient of variation of a region and a user defined coefficient of variation. For segmentation of optical imagery, the cartoon model assumes that the variability of digital number is constant over the entire image. Conversely, the texture model uses the variability among regions to segment the image. A used defined coefficient of variance is used to define the homogeneity degree for both models. SegSAR was implemented in IDL (Interactive Data Language), and uses ENVI tools to make the interface user friendly. The performance of SegSAR over radar imagery was conducted using quantitative and qualitative criteria over simulated data and qualitative criteria over airborne band X image. The performance over optical data was conducted using qualitative criteria over Landsat 7/ETM+ and Landsat 5/TM imagery. Both, the cartoon and the texture models, showed good results for both the radar and optical imagery, except for the not satisfactory results obtained when the texture model was applied over simulated radar imagery.

ASSUNTO(S)

segmentation radar sensoriamento remoto remote sensing image processing multi-nível radar processamento de imagens segmentação landsat imagens ópticas optical images multi-level

ACESSO AO ARTIGO

Documentos Relacionados