Analysis of capsulenets towards hyperspectral classification

The deep network features are being widely explored for improving the classification of remote sensing images. However, for hyperspectral datasets, the spectral features are found to be more significant as compared to their spatial counterparts. In this study, a deep learning framework is proposed for modelling the spectral features. Unlike the conventional strategies, the approach simultaneously optimizes both the feature extraction and the classification stages. In this approach, the spectral features derived from different levels of hierarchies, re-modelled as capsules, are used to label the given spectrum based on an iterative dynamic routing process. Consequently, unlike the regular convolutional architectures, here the relative locations of the spectral artefacts are also taken into consideration. Along with the margin loss, a spectral-angle-based reconstruction loss is also employed to facilitate proper regularization. Experiments over different standard datasets indicate that the proposed approach performs better when compared to the prominent approaches. Furthermore, in comparison with the former deep learning models, our approach is found to be less sensitive to the network parameters and achieves better accuracy even with lesser network depth.