LAUSR

This paper presents an approach for hyperspectral image classification using contextual sparse coefficients based on sparse representations. The main idea is that the sparse coefficients obtained through sparse representation-based modelling of the hyperspectral images contain discriminative characters which can be utilized for hyperspectral image processing tasks, such as classifications. Moreover, such discriminative features can be enhanced by incorporating the contextual information in the sparse coefficient domain. The proposed method starts with finding a representative spectral dictionary using the training data. Sparse coding is then applied to obtain the sparse coefficients. This is followed by a contextual transform performed in the sparse coefficient domain to incorporate the spatial information. Finally, the contextual sparse coefficients are used as feature vectors for training traditional classifiers, namely SVMs and kNN in our case. The experimental results conducted on a number of hyperspectral data sets confirm the effectiveness of our approach.