Abstract: When we apply the active learning method to the hyperspectral image classification, there always exist problems that the spectral-spatial features cannot be effectively utilized and the samples need to be manually labeled. In this work, we propose an active learning method combined with the convolutional neural networks to solve the above problems. Firstly, the neighborhood pixels of each pixel are extracted to form training samples. Then the spatial and spectral features of the samples are trained by the convolutional neural network and the data are preliminarily classified. Then, based on the spatial similarity and spectral similarity of the hyperspectral image, some of the unlabeled samples are labeled and added to the training set, which will further increase the accuracy of the classifier. Validated by Salinas image, Pavia University image, and Indian Pines image, the proposed method effectively improves the classification accuracy with less labeled samples.

Key words: hyperspectral image classification, active learning, convolutional neural network, spatial-spectral feature