Hyperspectral Image Classification Using Metric Learning in One-Dimensional Embedding Framework

Abstract

Hyperspectral image (HSI) classification has become an active research area in the remote sensing field. In order to construct a simple and reliable classifier, learning an adequate distance metric from a given HSI dataset is still a critical and challenging task in many HSI applications. In this paper, a novel distance metric learning (DML) framework based on 1-D manifold embedding (1DME), named DL1DME, is proposed for HSI classification. The 1DME framework was developed by using the recently developed smooth ordering technique. This framework enables us to elaborately exploit the benefits of DML in the development of the 1DME algorithm. The core of the state-of-the-art DML is to learn a Mahalanobis matrix from the given dataset that better describes the similarity between pixels. Largest margin nearest neighbors (LMNN) and information theoretic metric learning (ITML) are employed for the Mahalanobis matrix learning. Then, based on the affinity defined by the Mahalanobis matrix, the preclassifiers are constructed using the simple 1-D regularization on 1DME; and they predict the labels of the test data. By a voting rule, the pixels labeled in the same class by most of the preclassifiers are voted into the confidently predicted set, which are then merged with the current labeled set. The labeled set enlargement process is repeated if the original one has a very small size. The final classifier is then constructed in the 1DME framework again, but based on the enlarged labeled set. According to the aforementioned strategy, two novel DML-based 1DME classification algorithms, DL1DME-LMNN and DL1DME-ITML, are developed in this paper. Experimental results on three popular real-world HSIs demonstrate that the classification performance of the proposed DL1DME is superior to other most popular SSL methods in terms of classification accuracies.