关键词: 高光谱遥感影像分类, 空谱特征, 堆栈式稀疏自编码深度网络, 主动学习

Abstract: In order to avoid the problems that it is difficult and time-consuming to obtain enough labeled samples for hyperspectral image classification, while numerous unlabeled samples have not been used efficiently, and over-dependent on spectral information while ignoring spatial information, it proposes a spatial-spectral feature combined with active deep learning approach for hyperspectral image classification in this paper. Firstly, it extracts spatial-spectral feature by extracting a pixel’s small neighborhood pixels of a square as the spatial information and combines with the original spectral information on the basis of principal component analysis to the original image for dimension reduction. Secondly, it creates a sparse expression of raw hyperspectral image using sparse autoencoder. Then a deep neural network generating the deep features of raw data is built through learning stacked sparse autoencoder layer by layer. In addition, limited labeled samples are used to train the softmax classifier and fine-tune all the stacked sparse autoencoders in a supervised way. Finally, the most uncertain samples are selected to label with active learning algorithm and added to the training samples. Classification experiment on Pavia University image and Pavia Center image confirms that the proposed method can effectively improve the classification accuracy with less labeled samples comparing with traditional classification method.

Key words: hyperspectral image classification, spatial-spectral feature, stacked sparse autoencoders, active learning