Nonlinear data separation and fusion for multispectral image classification

The presented work deals with the problem of remote sensing data separation and fusion. Multispectral images are acquired from different bands. The collected radiances are the results of many reflections due to the land heterogeneity and the atmosphere. The mixture phenomenon is therefore nonlinear. This work aims to find an adequate nonlinear separation model based on Bayesian inferences. Sources are considered as gaussians and the nonlinearity is implemented by one hidden layer neuron network. The extracted sources have initially the same dimension as the observations. To reduce the Hughes phenomenon illness a dimension reduction algorithm will be proposed. We will select a subset of sources that describe efficiently the ground truth. The resulting source set will be called primary sources. After that, remain sources will be used to smooth the primary source classification results. The major goal of the presented work is to perform a powerful land characterization that describes the land more efficiently than observations Keywords— Remote sensing imaging; Hughes phenomenon; Source separation; Bayesian model; Dimension reduction; Data fusion; classification.