An efficient texture image segmentation algorithm based on the GMRF model for classification of remotely sensed imagery

Abstract

Texture analysis of remote sensing images based on classification of area units represented in image segments is usually more accurate than operating on an individual pixel basis. In this paper we suggest a two-step procedure to segment texture patterns in remotely sensed data. An image is first classified based on texture analysis using a multi-parameter and multi-scale technique. The intermediate results are then treated as initial segments for subsequent segmentation based on the Gaussian Markov random field (GMRF) model. The segmentation procedure seeks to merge pairs of segments with the minimum variance difference. Experiments using real data prove that the two-step procedure improves both computational efficiency and accuracy of texture classification.

Texture classification of Mediterranean land cover

“…These three components can be measured to provide information on spatial variation for use in land cover classification. A variety of texture measures is available ranging from simple statistics, such as standard deviation, to fractals (Parrinello and Vaughan, 2002), Markov random fields (Li and Gong, 2005), Gabor functions (Angelo and Haertel, 2003), wavelets (Carvalho et al, 2004;Myint et al, 2004), Local Binary Patterns (Lucieer et al, 2005;Petrou and Sevilla, 2006), statistics derived from the co-occurrence matrix and functions such as the variogram.…”

Learning sparse conditional random fields to select features for land development classification

“…Therefore, it is important to use contextual information in the form of spatial dependencies in the images Liu 2005, Zhong andWang 2007a, b). In fact, the issue of using contextual information has been of on-going interest in the remote sensing community (Kontoes and Rokos 1996, Tso and Olsen 2005, Daniels 2006, Li and Gong 2006, Luo et al 2006, Ferreira et al 2007, Tang et al 2007. Markov random fields (MRFs) are the classical probabilistic approaches for modelling the contextual information in the label image (Li 2001, Tso and Olsen 2005, Li and Gong 2006, Luo et al 2006, Tang et al 2007).…”

A New Multiscale Segmentation Algorithm on Texture Image in Wavelet Domain

“…Three principal approaches used to describe texture are based on feature, model and structure [2]. Feature-based Method is commonly used in some kind of neighborhood, choosing one or several statistical quantities in a fixed window to represent the local variation pattern, for example, the spatial gray-level cooccurrence matrix [3]、the average texture energy of fixed windows [4]、multi-parameter and multi-scale features [5], etc. Model-based methods exploit statistical model parameters to describe the local variation pattern, and choose model expression to combine these patterns together to represent texture, for example, the Gauss Markov Random Field (GMRF) [6] , Multi-Resolution Markov Random Field (MRMRF) [7] and the Hidden Markov Tree (HMT) [8].…”