Parameter Estimation for Bayesian Classification of Multispectral Data Refaat M Mohamed and Aly A Farag Computer Vision and Image Processing Laboratory University of Louisville, Louisville, KY 40292, USA {refaat, farag}@ cvip.uofl.edu www.cvip.uofl.edu Abstract. In this paper, we present two algorithms for estimating the parameters of a Bayes classifier for remote sensing multispectral data. The first algorithm uses the Support Vector Machines (SVM) as a multi-dimensional density estimator. This algorithm is a supervised one in the sense that it needs in advance, the specification of the number of classes and some training samples for each class. The second algorithm employs the Expectation Maximization (EM) algorithm, in an unsupervised way, for estimating the number of classes and the parameters of each class in the data set. Performance comparison of the presented algorithms shows that the SVM- based classifier outperforms those based on Gaussian-based and Parzen window algorithms. We also show that the EM based classifier provides comparable results to Gaussian- based and Parzen window-based while is an unsupervised. Key words: Bayes classification, density estimation, support vector machines (SVM), expectation maximization (EM), multispectral data. 1 Introduction Bayes classifier constitutes the basic setup for a large category of classifiers. To implement the bayes classifier, there are two many two parameters to be addressed [1]. The first parameter is the estimation of the class conditional probability for each class, and the second one is the priori probability of each class. In this paper, we address both of these parameters. Support Vector Machines (SVM) was developed to solve the classification problem, but recently have been extended to regression problems [2]. SVM have been shown to perform well for density estimation where the probability distribution function of the feature vector X can be inferred from a random sample Ð. SVM represent Ð by a few number of support vectors and the associated kernels [3]. This paper employs the SVM as a density estimator for multi-dimensional feature spaces, and uses this estimate in a Bayes classifier. SVM as a density estimator work in a supervised setup; the number of classes and a design data sample from each class need to be available. There are a number of unsupervised algorithms for parameter estimation for Bayes classifier, e.g. the well known k-means algorithms [4]. The k-means algorithm