Research On Object-oriented Classification From Hyperspectral Imagery

As the increase in spatial resolution of hyperspectral imagery, spectral heterogeneity in the same ground object is enhanced. The traditional pixel-based classification technologies can not extract the perfect ground object because of the Salt and Pepper in the results, which go against the sequential process. The object-oriented classification technologies aiming at high spatial resolution imagery classification conquers the localization of traditional technologies which provide a novel approach for the classification of hyperspectral imagery with high spatial resolution. This paper gave a summary of classification and feature extraction methods of hyperspectral imagery,several researches on object-oriented hyperspectral imagery classification were made. The major works implemented are listed as follows:1.Firstly, the hyperspectral RS technologies were analyzed and the pixel-based classification technologies were summed up. Then the object-oriented classification technologies were investigated, and the potential of object-oriented hyperspectral imagery classification was expatiated.2.Several methods of radiometric and geometrical correction were summed up, the essentiality of dimension reduction was given from the aspect of classification, some feature extraction methods were illustrated, certain approaches of intrinsic dimension estimate for high-dimensional data were researched, and the results validated the importance of intrinsic dimension estimate to dimension reduction for hyperspectral imagery.3.Object-oriented RS imagery classification technologies was lucubrated, parameter setting to multiscale segmentation, constructing of hierarchical structure, feature extraction from objects and fuzzy classification based on rules were analysed in detail, generic rules for object-oriented hyperspectral imagery classification were gained. From the experiments, it can be concluded that this technologies can improve the accuracy and interpretation of classification at the same time.4.Mean Shift and Quick Shift were used for hyperspectral imagery segment, which predigested the complicated operations of object-oriented classification.The result indicated that the accuracy was better via proposed methods than pixel-based approaches, and the efficiency was boosted evidently.