Non-linear methods in remotely sensed multispectral data classification

The aim of this research is to develop an effective approach being able to deal with the stochastic nature of remote sensing data. In order to achieve this objective it is necessary to structure the methodological knowledge in the area of data mining and reveal the most suitable methods for the prediction and decision support based on large amounts of multispectral data. The idea is to establish a framework by decomposing the task into functionality objectives and to allow the end-user to experiment with a set of classification methods and select the best methods for specific applications. As a first step, we compare our results from Bayesian classification based on non-parametric probability density estimates of the data to the results obtained from other classification methods. Tree scenarios are considered, making use of a small benchmark dataset, a larger dataset from Corine land cover project for Bulgaria and analyzing different features and feature selection methods. We show that the theoretically optimal Bayesian classification can also achieve optimal classification in practice and provides a realistic interpretation of the world where land cover classes intergrade gradually.