Spectral-spatial hyperspectral image classification with dual spatial ensemble learning

Abstract

ABSTRACT The spatial information in hyperspectral images (HSIs) plays a vital role in supervised classification tasks. However, although the introduction to spatial information can effectively improve the classification accuracy for representation-based supervised classifiers, it is difficult to avoid the expensive cost of time consumption. To solve this problem, a dual spatial ensemble learning (DSEL) method is proposed in this letter, which consists of the following key technologies: 1) first, the principal component analysis is applied to original hyperspectral data to extract the first three principal components (PCs). Then, the first three PCs are utilized as a base image of an over-segmentation method to cluster the HSI into many shape-adaptive superpixels. Finally, the raw HSI is guided by superpixels to construct new feature data with spatial information. 2) A representation-based classifier is used on the feature data to achieve initial classification results. Specifically, the results are converted into probability based on mathematical statistics. 3) A random walk algorithm is exploited to optimize the probabilities according to the spatial relationship between pixels. Experimental results on the Indian Pines and University of Pavia datasets demonstrated that the proposed DSEL method achieves superior performance compared with several competitive methods.