Research On Hyperspectral Image Classification Based On Improved Local Binary Model And Semi-supervised Learning

Hyperspectral images have pairs of continuous spectral bands,which contain abundant spectral information and spatial information of ground objects on the earth’s surface,so that some objects that cannot be identified by conventional remote sensing methods can be identified in hyperspectral images.However,the rich data information increases the difficulty of data processing and analysis,and there are problems such as difficulty in sample labeling.In order to improve the classification accuracy of hyperspectral images,this thesis conducts a progressive study,and first proposes a feature extraction method that improves the Local Binary Pattern(LBP)operator,which compresses the feature dimension without affecting the classification accuracy;Then,the unlabeled sample labeling method based on semi-supervised learning is studied to expand the labeled sample set;finally,the classification accuracy is further improved by improving the sparse representation model.The main content completed in this thesis are as follows:(1)Aiming at the problems of strong correlation between bands of hyperspectral images,redundant information,high data dimension,and complex processing,a feature extraction method based on improved LBP operator is proposed,which compresses the feature dimension without affecting the classification accuracy.It solves the problem of redundant feature extraction by conventional LBP operators,extracts texture features for hyperspectral images,and provides reliable information for subsequent further processing and classification.(2)Aiming at the problem of limited hyperspectral image label samples,a semi-supervised learning-based hyperspectral image sample labeling method is proposed.This method uses a sample screening strategy to find some samples worthy of labeling from a large number of unlabeled samples,and finally marks the filtered samples based on the neighborhood information and priority classifier to obtain the learned pseudo-labeled samples,so as to expand the labeled samples set.(3)Aiming at the redundancy problem of traditional sparse dictionary in the hyperspectral image classification model based on sparse representation,a nearest neighbor based sparse representation classification model is proposed.The model builds an exclusive dictionary for each sample to be classified,which avoids the complicated calculation process with the original redundant dictionary and reduces the interference of noise.(4)The classification effect of the proposed method on three datasets is verified by experiments.The experiments show that the proposed method has achieved good classification results on the Salinas Scene dataset and the Pavia University dataset.