Automatic framework for semi-supervised hyperspectral image classification using self-training with data editing

Abstract

In this paper, we propose a new semi-supervised classification algorithm called RDE_self-training, which is an automatic framework for classification of remotely sensed hyperspectral images. The algorithm exploits abundant unlabeled samples when the number of labeled samples is limited to learn an accurate classifier. Train the classifier iteratively on enlarged training set with data editing. Firstly, train a classifier with initial labeled samples and predict the unlabeled samples. Secondly, revise the labels of mislabeled samples according to nearest neighbor voting rule. Thirdly, select few samples ordered by high probability from the revised samples set, and filter the noisy samples to enlarge training set, then retrain the classifier to predict. Finally, revise the mislabeled samples according to nearest neighbor voting rule to obtain the final classification map. During the process of semi-supervised classification, the unlabeled samples are selected from the pool of candidates automatically without human effort. The effectiveness of the proposed approach is evaluated via experiments with real hyperspectral image collected by AVIRIS over the Indian Pines region, Indiana. The experimental results indicate that the proposed framework outperform state-of-the-art classification performance with unlabeled data added.