Semi-supervised Ensemble Learning For Hyperspectral Image Classification

Hyperspectral image(HSI)classification is a hot topic in the image processing field currently,the researchers focus on finding different technical methods to make computer to simulate human learning activities,analyze data automatically and learn images intel igently.In hyperspectral image,labeled instance is rare since it costs a lot of effort to obtain.Meanwhile unlabeled data is easy to collect relatively,but there has few ways to use them.In this context,semi-supervised learning theory receives growing attention,it uses unlabeled instance for model training,therefore classification accuracy can be improved effectively.Based on the analysis of the,how to combine ensemble learning with semi-supervised learning for further improvement the effectiveness of the proposed classification algorithm is a promising research direction.In this paper,three hyperspectral image classification methods are proposed.Applying these proposed methods on three hyperspectral data sets which are Indian Pines,University of Pavia and Salinas Scene,we conduct experiments and achieve satisfying classification accuracy.The main contributions can be summarized as follows:1.We develop a semi-supervised hyperspectral image classification method based on gradient boosting decision tree(GBDT)which combines active learning(AL)and semisupervised learning.In the model training process of the algorithm,not only these samples with highest confidence are added to the training set,but also the most controversial samples is selected to human experts to label and then the sample is used for expand the training set.At the same time,using the gradient boosting decision tree as a classifier and maximizing the performance of the classifier with small amount of labeled instance.2.We propose a semi-supervised HSI classification method based on ensemble diversity which design classifier from two aspects: explicit and implicit.First,improving the difference of the input samples by applying dictionary representation.Second,maximizing accuracy of base classifier on labeled data while maximizing diversity among them on unlabeled data.Experiments are conducted on three hyperspectral remote sensing data sets acquired by different sensors,and experimental results demonstrate the efficiency of the proposed method.3.We propose a novel HSI classification method based on gradient boosting decision tree feature extraction and semi-supervised learning.This method exploits different functions for unlabeled data selection,which is based on the evaluation of two criteria: uncertainty and diversity.This method aims at selecting a set of unlabeled samples that are easy to distinguish and representative,so as to improve the performance of models.At the same time,maximize the discriminative of feature based on GBDT.