A contextual classification system for remote sensing using a multivariate Gaussian MRF model

Abstract

We propose a spatial contextual classification system for remote sensing images. In the system the observed multispectral images are modeled with a multivariate Gaussian Markov Random Field (GMRF) model and the hidden classified image is modeled with another type of MRF model. The classification is carried out from the viewpoint of Maximum a Posteriori (MAP) estimation. One of the well-known problems of MAP estimation is its high computational complexity involved. One way to avoid this problem is a pixelwise classification that is successfully implemented on a computer with a clique-type block matrix notation of a multivariate GMRF local conditional density function (LCDF). The proposed system is applied to real remote sensing data. INTRODUCTIONAs a result of recent development in remote sensing technologies, quite high-dimensional multispectral image data are now available. To analyze these data, it is important to classify each pixel into an appropriate class, however the conventional pixelwise spectral-base classification methods sometimes bring about misclassifications because of not considering spatial correlations. Thus spatial contextual classification methods using both spatial and spectral information is desirable to ameliorate classification accuracy. One of the easiest way to implement a spatial contextual classification is to operate a smoothing spatial filter as a preprocessing or postprocessing, however it is difficult to find any physical justification of using such a filter. Accordingly we developed a classification system using stochastic models like the Markov Random Field (MRF) that enable us to utilize both spatial and spectral information simultaneously (see also [SI).Over the last decade MRF models have been utilized in the fields of image restoration, classification and segmentation. Some researchers discussed the mathematical properties of MRF models and many concrete energy functions were introduced. The Gaussian MRF (GMRF) model is one of a variety of MRF models and various properties of multivariate GMRF model were minutely studied. This paper is organized as follows: A two-component MRF system is proposed in Section 2. Then the parameter estimation methods are developed in Section 3. In Section 4, we describe the image classification criterion and the experimental results are given in Section 5.

Unsupervised segmentation of colored texture images by using multiple GMRF models and a hypothesis of merging primitives

“…The aforementioned methods dealt with conventional Markov fields. With actual segmentation, however, Gaussian distribution is more desirable in order to not impose restrictions on processed images; in so doing, obtaining relatively large areas through repetitive merger of small segments is preferable to seeking closed-edge areas [7,16]. Kaneko and Yodogawa [5] and Chellappa and colleagues [2] proved efficiency of Gaussian Markov fields (GMRF) for texture classification.…”

Combinação de modelos de campos aleatórios markovianos para classificação contextual de imagens multiespectrais

“…Tal modelagem foi originalmente proposta em (80)(81), mas além de ser restrita apenas a sistemas de vizinhança de primeira ordem (definidos pelos quatro vizinhos mais próximos), os autores utilizaram apenas o algoritmo ICM (determinístico) para aproximar o estimador MAP do campo de rótulos resultante. Nossa metodologia utiliza sistemas de vizinhança de ordens superiores (2 a e 3 a ordens), bem como a inclusão e combinação de algoritmos não determinísticos, como o MPM e o GSA na aproximação do estimador resultante.…”

Remote sensing image classification for forestry using MRF models and VQ method

“…Following the first part, the MRF model parameters are estimated for M known classes using the Maximum Likelihood (ML) method [4]. Afterward the sites in an unknown class are re-classified into M known classes using the MAP estimation.…”

“…In this paper we apply ML and MPL estimation methods for parameter estimation. Details on parameter estimation can be found in [4].…”

“…Although classification of water areas from the other is relatively easy, distinction between deciduous tree areas and coniferous (ever-green) tree areas is difficult because of the similarities of electromagnetic radiation patterns and the conventional pixelwise classification methods often cause isolated misclassifications. Recent researchers[ 11- [4] have disclosed that a contextual classification, that utilizes spatial and sometimes temporal information as well as spectral information, makes it possible to reduce such misclamifications. Markov Random Field (MRF) theory provides a convenient and consistent way of involving the contextual information in image models.…”