Research On Self-training Semi-supervised Classification For Hyperspectral Remote Sensing Image Based On Information Entropy

Hyperspectral remote sensing technology is a research hotspot, and has been a symbol of milestone in the modern remote sensing field. The automatic classification technology of hyperspectral image is the most important of hyperspectral remote sensing. In order to solve the practical problems of the automatic classification technology of hyperspectral image, e.g. high classification accuracy, stability and universality, the hyperspectral classification technology is studied in this dissertation. According to the analysis of the existing problems in the supervised and unsupervised method, the investigation is conducted about self-training semi-supervised theory and information entropy theory. The major works and contribution of this dissertation are listed as follows:(1) To solve the existing problems in traditional supervised classification method, e.g. high marked labels, quantity requirements, the low classification accuracy and serious misclassification, a new semi-supervised classification method of hyperspectral image is proposed based on Renyi entropy and multinomial logistic regression algorithm. Firstly, a small amount of labeled sample data is selected by using the multinomial logistic regression algorithm. The unknown sample categories information and probability information were predicted and outputted by the classifier. Secondly, the entropy of the unknown samples is calculated by Renyi entropy calculation method, and some unlabeled samples of maximum Renyi entropy are selected from the calculation data and added to the sample data. Finally, the classification of multinomial logistic regression is not iterated repeatedly for many times until the classification accuracy tends to be stable. According to the experimental results of five different hyperspectral remote sensing data, it can be found that the accuracy is improved about 1%-25.93% by comparing with the conventional supervised methods. The effect map of classification are superior to the traditional unsupervised and supervised methods.(2) Based on the theory of combining with posterior probability SVM algorithm and D-S evidence, a new semi-supervised classification method of hyperspectral image is further proposed. The speed of the self-training semi-supervised classification method is enhanced by using the rapid prediction ability of probability support vector algorithm, and the quality is improved by using the combination rules of evidence theory because of changing the unlabeled samples into labeled samples. According to the comparison experiments of the semi-supervised classification method based on combining with Renyi entropy and multinomial logistic regression algorithm, the analysis results show the accuracy is slightly improved to maximize 4.46%. The efficiency is greatly improved about 25.97%-59.70%.(3) Based on noise-estimation and minimum Renyi cross-entropy theory, a new band selection method of hyperspectral remote sensing image is proposes. The influence of dimension reduction on the efficiency and quality of self-training semi-supervised classification method is analyzed by using the new band selection method. Basically, the core idea of this algorithm has three steps. Firstly, the random noise in hyperspectral image is estimated by using a linear regression estimation method. Secondly, the band of the maximum spectral information is selected by using the Renyi cross entropy theory. Finally, the band of the maximum spectral information is combined by using Pearson correlation coefficient. The experimental results show that the proposed method can significantly improve the working efficiency of semi-supervised classification method about 7%-30% after dimension reduction by removing the redundant bands and preserving the maximum spectral information in hyperspectral data.