毕业生知识库

 

Target (land cover) classification for hyperspectral images is a hot topic in remote sensing image processing. It is an important and fundamental technique for developing hyperspectral image analyzing systems in various types of applications. Technically, the task of target classification for hyperspectral images is involved the following issues: band selection, feature representation and classifier construction. Although there are many thoughtful attempts in the past years, it is still a challenging problem due to the following three reasons. (1) The spectral resolution of hyperspectral images is usually high, and the images could be easily degraded by various kinds of noises, resulting in many outlier channels (bands). Employing all of the spectral bands to train a classifier could reduce the classification performance. (2) The spatial resolution of hyperspectral images is usually

low and many small targets belong to the same classes often locate in different spatial regions. It is laborious to label all of the small regions (especially for those targets spatially located in hundreds of thousands tiny regions) for training a desired classifier. (3) Given a small number of labeled samples with high dimensionality, “Hughes”phenomenon in training and prediction may occur, which could finally degrade the classification accuracy. Taking the above aspects into consideration, in this dissertation, we will focus on the following issues: band selection, multiple feature selection and classifier construction in terms of sparse learning, unsupervised learning and semi-supervised learning. Specifically, the main contributions are listed as follows:

(i) A band selection approach based on the structured sparse learning and semi-supervised learning is proposed. Specifically, the core idea is to consider the task of band selection as a problem of feature selection, which incorporates the structured sparse constraint and manifold regularization to improve the robustness and accuracy of band selection. On the one hand, in the level of spectral bands, a ℓ2,1 norm structured sparse constraint is incorporated into the model to improve its robustness to noise bands. On the other hand, in the level of spatial pixels, a trick of manifold regularization is employed to describe the local smoothness for spatial neighbors. As a result, the selected bands by the proposed method have better discriminative ability for land cover classification. Comparative experiments validate the effectiveness of the proposed method.

(ii) A multi-view unsupervised feature selection approach is proposed for multiple features of hyperspectral images. Specifically, the core idea is to map the spectral features and low-level visual features of targets into a low dimensional embedding representation using the trick of multi-view embedding. To achieve this goal, a ℓ2,1 structured sparse robust linear regression model is constructed to preserve the low dimensional embedding structures. As a result, the proposed model explores and exploits the complementary information contained in the multiple features. Meanwhile, the low dimensional embedding representation preserves latently the clustering and discriminative information, indicating the model is more accurate for feature selection. Comparative studies validate the effectiveness of the proposed method.

(iii) The task of land cover classification for hyperspectral images is often confronted with the problem of a small number of labeled samples with high dimensionality. A typical existing methodology for this kind of problem is the technique of graph-based label propagation that is able to utilize the large number of unlabeled samples in semi-supervised learning. However, the spatial resolution of hypersepctral remote sensing images is usually low, and the land covers are usually spatially located in many small regions in the images. Technically, these small regions would be smoothed away during label propagation, resulting in a poor classification. To solve this drawback, we proposed a semi-supervised classification approach, which can automatically produce labeled samples. First, a classifier is trained with the small number of labeled samples, and the trained classifier is then employed to classify the hyperspectral images in pixel-level. Based on the pixel-level classification results, new seeds (namely pixels) are identified

and produced as newly-labeled ones under a criterion of label consistency in local regions. Accordingly, the label probabilities of these new seeds together with the previous ones offered in advance are propagated on a graph, which is constructed to preserve the local edges. Finally, the label propagation problem is addressed in terms of semi-supervised classification and formulated as an energy minimization problem, which can be solved via a sparse linear system. Comparative studies show that the proposed

method is superior to the existing methods on classification accuracy.