Hyperspectral image classification using k-sparse denoising autoencoder and spectral–restricted spatial characteristics

Abstract

Hyperspectral images (HSIs) have both spectral and spatial characteristics that possess considerable information. This paper proposes a novel k-sparse denoising autoencoder (KDAE) with a softmax classifier for HSI classification. Based on the stack-type autoencoder, KDAE adopts k-sparsity and random noise, employs the dropout method at the hidden layers, and finally classifies HSIs through the softmax classifier. Moreover, an operation referred to as restricted spatial information (RSI) is conducted to obtain the spatial information of the HSI. The proposed method extracts features by KDAE combined with RSI, effectively taking the spectral and spatial information into account. The obtained features are finally fed into the softmax classifier for classification. The experimental results obtained on two benchmark hyperspectral datasets indicate that the proposed method achieves state-of-the-art performance compared with other competing methods.