An efficient semi-supervised classification approach for hyperspectral imagery

In this paper, an efficient semi-supervised support vector machine (SVM) with segmentation-based ensemble (S²SVMSE) algorithm is proposed for hyperspectral image classification. The algorithm utilizes spatial information extracted by a segmentation algorithm for unlabeled sample selection. The unlabeled samples that are the most similar to the labeled ones are found and the candidate set of unlabeled samples to be chosen is enlarged to the corresponding image segments. To ensure the finally selected unlabeled samples be spatially widely distributed and less correlated, random selection is conducted with the flexibility of the number of unlabeled samples actually participating in semi-supervised learning. Classification is also refined through a spectral-spatial feature ensemble technique. The proposed method with very limited labeled training samples is evaluated via experiments with two real hyperspectral images, where it outperforms the fully supervised SVM and the semi-supervised version without spectral-spatial ensemble.