Hyperspectral Image Classification Based On Semi-supervised Collaboration-training Algorithm

The classification of hyperspectral remote sensing images has been a hot research topic in the field of image processing for a long time,which is of great significance for land cover research,environmental monitoring and other fields.Traditional supervised classification methods are usually limited to the shortage of labeled samples that are also difficult to obtain,so new methods need to be developed and applied.Semi-supervised collaboration-training(Co-training)algorithm can expand the training sample set through the base-classifiers provide high confidence "pseudo-labeled" samples to each other,which minimizes the demand for labeled samples and has high application value in hyperspectral classification.However,traditional Co-training algorithms are faced with the problems that the initial classifier is easily affected by noisy data and the redundancy caused by the inter-selection of samples which result in accumulating error information in training stage.In order to overcome the common shortage of labeled samples of hyperspectral image classification and the existing problems of Co-training algorithm,this paper fully considers the data characteristics of hyperspectral images and the applicable scenarios of Co-training algorithm,and then proposes two semi-supervised hyperspectral classification methods based on the modified standard Co-training algorithm from two perspectives of prior distribution of data and high confidence criterion of candidate unlabeled samples,which are expected to further improve the classification accuracy of hyperspectral image.The tests results on Pavia Centre,Salinas and Kennedy Space Centre hyperspectral data sets are satisfactory,and the feasibility of the proposed method is verified.The main work is summarized as follows:(1)In this paper,the Co-training algorithms are applied to the classification of hyperspectral images,and the theoretical framework is analyzed and summarized.Through the way of combining theory and practice,the typical Co-training algorithms including standard Co-training and Tri-training are used to classify three open hyperspectral data sets.The results show that Co-training can achieve good classification results even the number of training samples is small;(2)Considering the influence of noisy samples on the Co-training algorithm,this paper uses local outlier factor(LOF)outlier detection algorithm to mine the prior distribution information of unlabeled samples by detecting global and local outliers,and then pick out high-quality samples that have concentrated information which are then labeled by the base-classifier.Thus this paper proposes a hyperspectral Co-training classification method based on local outlier factor(LOF)outlier detection which called LOF-Cotraining.This method avoids redundant computation for a large number of unlabeled samples that helps accelerate the convergence speed and reduces the introduction of noise samples.The experiment results on three open hyperspectral datasets show that the overall classification accuracy of LOF-Cotraining algorithm is at least 2.0%higher than that of comparison algorithms,and it can obtain high precision even there are fewer labeled samples;(3)Aiming at the problem of low accuracy of labeling "pseudo-labeled" samples in traditional Co-training algorithm,this paper adopt active learning for setting the criterion of additional unlabeled samples,and proposes a semi-supervised Co-training classification method based on active learning label checking named DC(Double Checker)-Cotraining.In this method,active learning is used to train a label checker for each base-classifier of Co-training.In the process of sample selection,the prediction labels given by the base-classifier are discriminated twice,so as to determine the "pseudo-labeled" samples with high confidence to be added to the training sample set which help improve the labeling accuracy of samples.Experiments are carried out on three open hyperspectral datasets,the results indicate that the overall classification accuracy of DC-Cotraining algorithm can be increased by up to 7.8%compared with the standard Co-training algorithm in the same case.