A New Sparse Source Separation-Based Classification Approach

In many geoscience applications, we have to convert remotely sensed images to ground cover maps. Numerous approaches to extract ground cover information have been developed. Recently, blind source separation (BSS) of remote-sensing data has received significant attention due to its suitability to recover sources when no information is available about the scanned zone, hence the term blind. In the remote-sensing context, associating each source to a significant land cover theme is difficult and constitutes the real challenge of this paper. Many authors have pointed out that BSS is overwhelmingly a question of contrast and diversity. This reasoning motivates this work which takes advantage of both decorrelation and sparsity to propose a two-level novel approach to separate our different land covers called sources. The first separation stage is based on second-order statistics or decorrelation. It gives a suitable representation of the remote-sensing images. However, decorrelation is a natural way of differentiating statistically between sources but is unable to identify and extract finer features with physical meaning. The aim of the second separation stage is to overcome this problem by an increasingly popular and powerful assumption which is the sparse representation. The last leads to good separation because most of the energy in the defined basis, at any time instant, belongs to a single source. This allows the extraction of physical features and the capture of image essential structures. The innovative aspect of this study concerns the development of a new image classification approach that integrates the BSS at the feature extraction level to provide the most relevant sources from remotely sensed images. It can be viewed as an unsupervised classification method. The second-order separation process is used as a preprocessing step to remove the interband correlation which sometimes brings ill effect to image classification. However, the second-order process is unable to uncover the underlying sources. The basic idea behind our approach is that heterogeneous multichannel data provide sparse spectral signatures in addition to sparse spatial morphologies in specified dictionaries. Hence, sparse modeling can be used to disentangle the land covers from observed mixtures. From the sparse representation, the data space is transformed into a feature space composed of mutually exclusive classes. Finally, we will merge these classes at the decision level in order to enhance the semantic capability and the reliability of land cover classification. The effectiveness of the proposed approach was demonstrated by operating two experiments to study respectively the source separation and the image classification capability of the developed approach. The different results on remote-sensing images illustrate the good performance of the new sparse approach and its robustness to noise. These experiments show that the sparse representation enhances the separation quality and allows extracting more easily the essential structures of the scanned zone. The proposed approach offers an interesting solution to the classification process with limited knowledge of ground truth.