Research On Classification Technology For Soil Elements And Man-made Objects Of Hyperspectral Imagery

Soil elements and man-made objects play an important role in geographic space information guarantee of military surveying and mapping. Hyperspectral imagery containing abundant spectral information provides a new research direction for detecting features attribute information and achieving features fine classification. In order to improve the classification accuracy of soil elements and man-made objects in hyperspectral imagery, this paper has carried out the thorough research on classification technologies for soil elements and man-made objects of hyperspectral imagery, respectively from the classification recognition based on spectral features and classification identification based on statistical characteristics, on the basis of summaring classification methods for soil elements and man-made objects of hyperspectral imagery. The main work and achievements are as follows:(l)The research status of classification methods for soil elements and man-made objects of hyperspectral imagery is summarized; the classification systems and their characteristics of different disciplines for soil elements and man-made objects are analyzed and summarized; the geographic distribution of soil elements is discribed; this paper analyzes spectral characteristic of soil elements and man-made objects, summarizes the traditional spectral similarity metrics, and expounds the main ideas of the classification method based on spectral characteristics.(2)In classification method based on spectral characteristics, according to the spectral characteristics of soil elements and man-made objects, Spectral Angle Cosine-Correlation Coefficient mrtric is formed combining Spectral Angle Cosine measure with Correlation Coefficient measure. A spectral matching classification method based on Spectral Angle Cosine-Correlation Coefficient mrtric is proposed combining with the continuum removed algorithm. The classification experiments show that higher classification accuracy for soil elements and man-made objects can be reached with Spectral Angle Cosine-Correlation Coefficient metric.(3)In classification method based on statistical characteristics, fully use of statistical characteristics of imagery high-dimensional data can improve the classification accuracy of soil elements and man-made objects. Kernel function methods and greedy forward algorithm are introduced to the logistic regression model to build a stable Import Vector Machine (IVM) classifier, which can be directly output posteriori probability information. The classification experiments show that this method can achieve classification accuracy equivalent to that of Support Vector Machine (SVM), with better stability and stronger sparsity.(4)There is a strong correlation between adjacent pixels of soil elements and man-made objects. Markov Random Field (MRF) model can be used to deal with the IVM classification results, which can eliminate the category noise in the homogenous area of the classification results combining with the space information. The ICM method is used to calculate the optimal value of the combined energy function, which is formed combining spectral energy function and space energy function. The comparison between the former results and the treated ones shows that Markov Random Field model is beneficial to the improvement of the imagery classification accuracy.