Improvements to Expectation-Maximization approach for unsupervised classification of remote sensing data
Thales Sehn Körting, Luciano Vieira Dutra, Leila María García Fonseca, Guaraci J. Erthal, Felipe Castro da Silva · Biblioteca Digital da Memória Científica do INPE (National Institute for Space Research) · 2007
Abstract. In statistical pattern recognition, mixture models allow a formal approach to unsupervised learning. This work aims to present a modification of the Expectation-Maximization clustering method applied to remote sensing images. The stability of its convergence has been increased by supplying the results of the well-known K-Means algorithm, as seed points. Hence, the accuracy has been improved by applying cluster validity measures to each configuration, varying the initial number of clusters. High-resolution urban scenes has been tested, and we show a comparison to supervised classification results. Performance tests were also realized, showing the improvements of our proposal, in comparison to the original one. 1.