This article proposes a non-supervised segmentation approach for multi-sensor remote-sensing data. Attention is focused on the phase of automatic training in order to retrieve the image class parameters, needed in the successive parametric segmentation. Since the clustering of the whole data set is in general not possible due to the computational load involved, sampling is needed that allows one to estimate the distribution of the classes in the feature space. However, on the one hand the sampling of single pixels may effect strongly the correct estimation of the class distributions due to noise, while, on the other hand, simply taking the mean value of a window around the sample may have too storing a filtering effect. The proposed algorithm exploits the spatial interaction between the pixels in the image, taking carefully into account the local image content of each sample of the observed scene. A Bayesian network estimates for each candidate sample the most appropriate neighborhood, looking for connected components and thus for pixels that are likely to partake of the same class. From the selected neighborhood a mean value of the feature vector is computed that is to represent the sample, thus taking into account the local morphologic information. In this way the estimated class distributions in the feature space form a more robust representation of the true classes, thus bearing advantages to the parameter estimation and to the final segmentation. The image class models obtained by the proposed training step are used as input to a Markov random field (MRF) segmentation approach. Results presented show that a better separation of the natural classes is possible when using in a more careful fashion the local image content. Numerical results based on synthetic images show that the accuracy of the MRF segmentation approach improves from 72 percent to 96 percent.