Semi-Supervized Hyperspectral Image Classification using Virtually Labeled Training Data based on the Improved Gaussian Mixture Clustering

摘要

This paper presents a semi-supervised approach for classifying hyperspectral remote sensing imagery using an improved clustering for training data selection. This approach is a development and an adaptation for hyperspectral imagery of the recent semi-supervised classification method proposed by the same authors for multispectral images. The present model follows the below steps: (a) dimensionality reduction of the hyperspectral pixels by Principal Component Analysis (PCA); (b) clustering in the low-dimensional space obtained after PCA processing using the Gaussian mixture model (GMM) with expectation-maximization (EM), which includes a K-means initialization phase; (c) selection, according to pixels' virtual labels given by clustering, of the training set; (d) inclusion in the training set of “M” pixels per class with genuine labels; (e) training of a semi-supervised SVM classifier (S3VM); (f) S3VM classification. For experiments, we have used the Pavia University hyperspectral image dataset. The experimental results have shown a significant performance improvement of the proposed semi-supervised classifier compared to the supervised SVM.