Semi-supervised local feature extraction of hyperspectral images over urban areas

We propose a novel Semi Supervised Local Embedding (SSLE) method for feature extraction from hyperspectral data. The proposed method combines a supervised method (Linear Discriminant Analysis (LDA)) and an unsupervised method (Local Linear Embedding (LLE)). The underlying idea is to get the Principal Components (PC) from the original data and input training samples from the principal components into LLE, LDA and into our proposed SSLE algorithm. Thereafter, Support Vetctor Machine (SVM) was used for classification. The overall accuracy of this new algorithm is then compared with other existing semi-supervised algorithms. Experiments on hyperspectral image show the efficacy of the proposed algorithm.