Expectation-Maximization × Self-Organizing Maps for Image classification Thales Sehn Korting, Leila Maria Garcia Fonseca National Institute for Space Research – INPE/DPI, S˜ ao Jos´ e dos Campos, Brazil {tkorting,leila}@dpi.inpe.br Fernando Lucas Bac ¸˜ ao Universidade Nova de Lisboa – UNL/ISEGI, Lisboa, Portugal bacao@isegi.unl.pt Abstract To deal with the huge volume of information provided by remote sensing satellites, which produce images used for agriculture monitoring, urban planning, deforestation de- tection and so on, several algorithms for image classifica- tion have been proposed in the literature. This article com- pares two approaches, called Expectation-Maximization (EM) and Self-Organizing Maps (SOM) applied to unsu- pervised image classification, i.e. data clustering without direct intervention of specialist guidance. Remote sensing images are used to test both algorithms, and results are shown concerning visual quality, matching rate and pro- cessing time. 1 Introduction The huge volume of information provided by remote sensing satellites is constantly growing, therefore the de- mand for algorithms which are able to deal with such data and produce valid results is also increasing. Satellites with different ground resolutions produce different kinds of im- ages, each of which with purposes, such as agriculture mon- itoring, urban planning, deforestation detection, etc. To deal with all this information, several approaches for image classification have been proposed on the literature. In the pattern recognition literature they are divided into two main types: supervised and unsupervised classification. The unsupervised techniques perform data clustering with- out direct intervention of specialist guidance. In this article we test two unsupervised approaches for remote sensing im- age classification: the Expectation-Maximization (EM) al- gorithm is compared with the Self-Organizing Maps (SOM) approach. Many methods have been developed to deal with image classification problems. The work of [2] proposed a frame- work of four kernel-based techniques for hyperspectral im- age classification using Support Vector Machines (SVM). Neural networks have also been used to perform classifi- cation, as in [12], which performs image segmentation to extract object regions, then submitted to classification us- ing shape and texture attributes. Other methods for re- mote sensing image classification include Gaussian Mixture Models through the EM method [13, 9], Self-Organizing Maps [7, 18, 16], Fuzzy sets and their combination with neural networks [14, 5], etc. In Section 2 we describe the EM and SOM algorithms applied to unsupervised image classification. In Section 3 we apply these algorithms to some images and compare the results to reference images created by a specialist. In the conclusion section we highlight the strong points of each algorithm, and make recommendations on their application. 2 The algorithms This section describes EM and SOM techniques ap- plied to unsupervised image classification. Both algorithms have been developed in the same C++ Library called Ter- raLib [1], available at http://www.terralib.org/, so that processing time for basic instructions will be the same for the two algorithms. 2.1 Expectation-Maximization Approach In statistical pattern recognition, mixture models allow a formal approach to unsupervised learning (i.e. cluster- ing) [4]. A standard method to fit finite mixture models to observed data is the EM algorithm, first proposed by [3]. EM is an iterative procedure which converges to a (local)