A Study Of Semi-supervised Learning On Hyperspectral Image Classification

Hyperspectral images(HSIs) are characterized by the high spectral and spatial resolution available, which allows capturing fine details of the spectral characteristics of materials in a wide range of applications. However, it also brings the challenges for HSI classification: 1) large number of spectral channels; 2) small number of labeled samples available; and 3) existence of noise and redundancy. These observations have fostered the use of feature mining and semi-supervised learning(SSL) techniques to address these issues. In this dissertation, we focus on SSL for HSI classification, which aims at addressing three important issues: design of robust semi-supervised HSI classification method, learning sparse discriminative features for HSI classification, and band selection for HSI. The main contributions of this dissertation are as follows:1. We propose a novel semi-supervised method based on a modified co-training process with spectral and spatial views for HSI classification. The proposed method exploits spectral-spatial two-views for co-training, which considers both the spectral and spatial information to improve the classification performance. Then a modified co-training process with a new sample selection scheme is presented, which can effectively improve the co-training performance especially when there are extremely limited labeled samples available.2. We propose a novel semi-supervised dictionary learning approach for HSI classification. The proposed method exploits both the labeled and unlabeled samples to learn the dictionary for reconstruction representation, and simultaneously optimizes the classifier parameters for classification. Therefore, it can learn a discriminative dictionary with improved generalization performance, and also accomplish the classification simultaneously. The classification is expected to be optimal since the classifier parameters are learned and optimized along with the sparse representation and dictionary.3. We propose a novel semi-supervised joint sparsity regularization based band selection method for HSIs. The proposed method is based on the linear regression model with the joint sparse norm regularization on the regression coefficient matrix to enforce selecting consistent bands for all the classes for classification. To improve the generalization ability, we also add the manifold(graph) regularization term in the proposed supervised model, which can effectively incorporate the discriminative information and the smoothness prior in a compact semi-supervised model. For band selection, only the maximum 2l-norms of the columns in the regression coefficient matrix need to be computed and the corresponding indices are selected, which is convenient for the user to select the specified number of bands.