Semi-supervised Hyperspectral Image Classification Based On Multinomial Logistic Regression And Neighborhood Information

The hyperspectral remote sensing technology has been widely applied in earth observation. However, the dimension of data is high and obtaining the initial training sample is a time-consuming process which leads to the problem of information redundancy; Multinomial logical regression classifier is with some advantages in high dimensional image data, but the performance of the classifier is mainly affected by the regression parameters, which makes the parameter optimization problem especially prominent. In this paper, semi-supervised classification and multinomial logistic regression classifier are applied to hyperspectral remote sensing image interpretation. In view of the existing problems, some new algorithms are proposed. The main contents are as follows:(1) According to the question that newton algorithm was often faced wtih time consuming and low accuracy in the process of regression parameters calculation, an improved method called DFP quasi-newton was proposed to improve the computational efficiency. The second order Hessian matrix is replaced by secant method in newton method and using DFP to improve the performance. The result shows that it has a big improvement in accuracy compare with the traditional multinomial logistic regression method.(2) In the process of semi-supervised classification, labeled samples directly determined the degree of improvement on the performance of the classifier. In view of the commonly used sample selection methods, a new improved algorithm was proposed. This algorithm considers the multi class problem. First effective classes were selected through the given threshold. After that, the relativity value of various effective classes was measured through variance. Finally the unlabeled samples with maximal values were labeled. Experiments show that: the proposed method can obtain better performance compared with original methods.(3) In the process of semi-supervised classification, label samples directly determine the final classification effect. The wrong sample label will not improve the results and lead to worse results. This paper presents a new algorithm that neighborhood information and classifier combination were used. First a circle search area was built. After that the training samples appearing in the search area was find out and structure a possible set. Finally the classifier results was compared with the possible set to confirm the final label. Experiments show that: compared with the original algorithm, this proposed algorithm was with obvious improvement.