Dimensionality-Varied Convolutional Neural Network for Hyperspectral Image Classification With Small-Sized Labeled Samples

摘要

Recently, convolutional neural network (CNN) achieves outstanding performance for hyperspectral image (HSI) classification. However, CNN requires large-sized labeled samples for training and tremendous computations for 3-D spectral-spatial feature extraction. In practice, most HSI data are unlabeled and the availability of samples for training is limited. In order to improve the classification performance with a simplified structure, a dimensionality-varied convolutional neural network (DV-CNN) is proposed. DV- CNN reduces the complexity and calculations by changing the dimensionalities of feature maps and improves the classification accuracy of small-sized labeled HSI by deep network. The experiments are compared with some state-of-the-art methods on Indian pines and Pavia University data sets. The experimental results demonstrate that the proposed DV-CNN outperforms other methods and achieves better classification performance.