Spectral-Spatial Active Learning with Superpixel Profile for Classification of Hyperspectral Images

摘要

Hyperspectral image (HSI) classification has to deal with scarcity of training samples. Active learning (AL) methods are employed in the literature for generating informative training samples. Most of the existing AL methods are based on spectral values alone. In this paper we propose a spectral-spatial AL model for classification of HSI having limited training samples. In the proposed model first the spectral and spatial information of the HSI is integrated by constructing an extended superpixel profile (ESPP). To this end, the dimension of HSI is reduced using principal component analysis and a superpixel profile (SPP) is constructed for each component image. The SPP is constructed by concatenating the component image with the results obtained by applying the simple linear iterative clustering technique and replacing with average values of superpixel considering a sequence of superpixel thresholds. The pixels are replaced with the ESPP features. Next a query function based on uncertainty, diversity, cluster-assumption and their combination are applied iteratively to select batch of most informative samples for including in training set. Experiments are conducted on two real HSI data sets in which the proposed model is compared with the models based on spectral values alone and the spectral-spatial model based on extended attribute profile. The AL methods in the proposed model has outperformed all the state-of-the-art AL methods.