A deep manifold learning approach for spatial-spectral classification with limited labeled training samples

One major challenge of designing deep learning systems for hyperspectral data classification is the lack of labeled training samples. Inspired by recent manifold learning researches, this paper presents a novel Locality Preserving Convolutional Network to address this challenge. The proposed method invents a semi-supervised locality-preserving regularization operation, and inserts a new layer in the three-dimensional convolutional neural network for end-to-end spatial-spectral classification. The benefits are three-fold. First, by using unlabeled training samples which are more easily available, the proposed method reduces the number of labeled samples required for training a deep learning model; Second, the proposed method incorporates the intrinsic geographical correlation among nearby samples into the extracted features, which prevents it from losing accuracy when only limited labeled samples are available; Third, with a three-dimensional architecture, the proposed method can extract the spatial and spectral features simultaneously from the hyperspectral data for classification. A gradient-decent based approach is deployed to train the whole network in a unified way. Experiments over different benchmarks show that, the proposed method relieves the Hughes phenomenon for deep learning, and achieves competitively high classification accuracy compared to other state-of-the-art approaches. (C) 2018 Elsevier B.V. All rights reserved.