Pre-trained Classification of Hyperspectral Images Using Denoising Autoencoders and Joint Features

Abstract

Hyperspectral image classification in remote sensing discipline aims to analyze scene properties of the environment captured from earth observing satellites of aircrafts. To comprehend this aim common linear methods like principal component analysis and linear discriminant analysis fail to model the nonlinear structures of data. To learn feature representations on large-scale high-dimensional data, deep learning methods have been applied successfully. We utilize a deep neural network for both feature extraction and then classification based on unsupervised pre-training using stacked denoising autoencoder method and supervised fine-tuning using logistic regression on top. This paper both exploit joint representation, namely spectral-spatial information of hyperspectral images to pre-train classification capturing the most salient features. Besides that, since extracting sparse features might improve the discrimination, rectified linear unit (ReLU) is used as activation function in encoders to extract high-level sparse features. The results show in our experiments that this model achieves the higher classification accuracy than other evaluation methods, and excels classical classifiers namely support vector machines and random forests.