Unsupervised Multi-manifold Classification of Hyperspectral Remote Sensing Images with Contractive Autoencoder

Abstract

Unsupervised classification is a crucial step in remote sensing hyperspectral image analysis where producing training labelled data is a laborious task. Hyperspectral imagery is basically of high-dimensions and indeed dimensionality reduction is considered a vital step in its preprocessing chain. A majority of conventional dimensionality reduction techniques rely on single global manifold assumptions and they can not handle data coming from a multi-manifold structure. In this paper, the unsupervised classification of hyperspectral imaging is addressed through a multi-manifold learning framework. To this end, this paper proposes a Contractive Autoencoder based multi-manifold spectral clustering algorithm for unsupervised classification of hyperspectral imagery. The proposed algorithm follows the same outline as the general multi-manifold clustering but exploits contractive autoencoder for tangent space estimation. We evaluate the proposed algorithm with two benchmark hyperspectral datasets, Salinas and Pavia Center Scene. The experimental results show the improvements made by the proposed method with respect to the conventional multi-manifold clustering based on local PCA and the basic autoencoder.