The land cover classification is an important task in geoscience applications. Many methods and implementations are based on multispectral data processing. The presented work aims to benefit from the nonlinear source separation process to enhance land cover identification. The source separation technique aims to provide underlying images and to compensate the mixing process. Nonlinear separation is more realistic due to multiple distortions occurring on the radiance path from soil to sensors. The presented paper addresses pattern recognition for remote sensing and proposes a framework based on feature extraction and decisional fusion. The first stage performs a nonlinear separation model based on Bayesian inferences. Nonlinearity is approximated by a multilayer neuron network. The separation process updates knowledge about unknown sources and model parameters iteratively. The second stage performs feature extraction. Based on a decisional fusion, the third stage realizes a classification process. This fusion scheme assigns, first, a suitable feature to each source/pattern based on the learning data set. Second, a majority vote determines the final label. Experimentation results demonstrate that the proposed fusion method enhances the recognition accuracy and represents a powerful tool for land identification.