CROSS-DOMAIN EXTREME LEARNING MACHINE FOR CLASSIFICATION OF HYPERSPECTRAL IMAGES

Cross domain extreme learning machine (CDELM) is an unsupervised domain adaptation algorithm. It achieves domain adaptation by minimizing the classification loss on source labeled data and utilizing the maximum mean discrepancy strategy. We apply this algorithm for classification of hyperspectral images, and improve it by introducing a transformation on source features, so that the target data and transformed source data can better share the same output weights in ELM network. The enhanced CDELM is denoted as ECDELM in this paper. The experimental results using Hyperion multi-temporal remote sensing images demonstrated the effectiveness of the improvement.