Spectral-Texture Feature Extraction And Multi-Classifier Ensemble For Hyperspectral Imagery

Hyperspectral remote sensing has been widely used in urban planning, land using and change detection, disaster monitoring, geology assessment, agriculture and forest investigation; it is an urgent problem that to improve processing efficiency and classification accuracy for sophomore applications. At present, different types of hyperspectral sensors have provided various hyperspectral remotely sensed imagery data, so effective processing algorithms are needed for more applications. Also, for remote sensing, the classification accuracy is mostly depend on the performance of classifiers; multi-classifier ensemble can get more higher classification accuracy than single classifier because they can provide complementary information for classification, it is the technology support for improving hyperspectral imagery classification accuracy.It should be noted that most available hyperspectral data processing techniques and classification algorithms are focused on analyzing the data without incorporating spatially adjacent information especially texture information on the data. In fact, there is not only spectral information, but abundant texture information in hyperspectral remotely sensed imagery. In certain applications, however, the incorporation of spatial and spectral information is mandatory to achieve sufficiently classification accurate results for hyperspectral imagery due to its spectral-spatial characteristics. Also, the existed methods for multi-classifier ensemble are mostly emphasized on ensemble algorithm for spectral information, and the algorithm combined texture with spectral information is still an open topic especially for hyperspectral remote sensing.In this paper, the foundation theory of hyperspectral remote sensing is introduced, and then hyperspectral band selection, feature extraction, texture description and extraction algorithms, and multi-classifier ensemble technology are investigated, a new multi-classifier ensemble algorithm which combined spectral feature with texture features is proposed. The details are as follows.(I) Band selection:the supervised and non-supervised hyperspectral band selection algorithms which based on orthogonal projection divergence (OPD) and adaptive affinity prorogation (AAP) clustering are presented, respectively. For OPD-based band selection method, it aims to discriminate the interesting objects from background and noise information, maximize the spectral similarity between different spectral vectors by projection the original data to feature space. At the same time, the sequential floating forward search (SFFS) is used during band selecting to reduce computation complex. For AAP-based on band selection, the "exemplar" numbers determinant algorithm and bisection method is addressed to improve AP procedure, and the relations between selected "exemplar" numbers and "preferences" are established. This paper presents the novel band selection algorithms for hyperspectral dimensionality reduction, and provides new strategies to improve classification accuracy for hyperspectral applications.(2) Feature extraction:the new supervised and semi-supervised hyperspectral feature extraction algorithms which based on the modified k-means clustering are put forward. For hyperspectral clustering, widely used k-means clustering is applied, and two novel initialization methods which based on OPD and similarity nonsupervised band selection are provided; they can separate the interesting target signatures from undesired signatures and background information, and can get the suitable initial seeds for k-means clustering. In addition, a new cardinality estimation index which maximizes the distance ratio between intra-cluster distance and inter-cluster distance (RICD) is presented; it is used as a tool to estimate the numbers of clusters in k-means for hyperspectral data. Moreover, the proposed methods are compared with other similar methods in computation complex and statistical significance testing.(3) Texture description and extraction:Traditional GLCM is extended into high dimensional space for hyperspectral imagery; the new term of hyperspectral volume texture and new texture model V-GLCM are proposed based on GLCM; the novel hyperspectral imagery texture description and extraction method which based on V-GLCM algorithm is designed. Hyperspectral imagery volume texture can be seen as the real function of different representation features of object points which in two-dimensional projection space in the spectrum space, and the collection values of the point sets constitutes of a volume texture. Based on the above analysis, the theory of window size analysis for V-GLCM is discussed and the semi-variogram function is chosen as the analysis tool. In the experiments, it has proved that our method performs better than GLCM.(4) Multi-classifier ensemble based on spectral and texture features:Firstly, the clustering subspace ensemble(CSE) method which based on spectral subspace clustering for hyperspectral imagery is proposed; for CSE. all the data can be partitioned into some groups by clustering analysis algorithm, and then a classifier is applied to each group of bands and the final output will be the fused result of multiple classifiers; different from other band grouping techniques(such as spectral correlation coefficient matrix), it allows nonadjacent bands to be clustered together, which may have high spectral correlation as well. Also, in this way, the number of training samples required is greatly reduced since the data dimensionality in each band group is much smaller than the dimensionality of the original data. Secondly, a new multi-classifier ensemble strategy which combined spectral data and texture data based on extracted features for hyperspectral imagery classification is investaged. In this method, it not only used discriminate characteristics in spectral space, but also used spatial information especially texture features in spatial space, the experiment results has demonstrated that it can improve classification accuracy for hyperspectral imagery.(5) Case study and applications analysis:The first case is endmember extraction and mineral recognition, and it is the most important application for hyperspectral remote image data. The second case is land use and land cover (LULC) classification in urban area, and LULC patterns is critical to urban monitoring. The proposed feature selection and extraction algorithm, multi-classifier ensemble method are applied to the two cases. by the experiments, it proved the efficiency and usability of the proposed methods.