A Cross-Domain Semi-Supervised Zero-Shot Learning Model for the Classification of Hyperspectral Images

The cost of obtaining spectral information has dropped significantly due to the remarkable progress in hyperspectral imagery. However, annotating the hyperspectral samples remains a labour-intensive and time-consuming behaviour. Given the challenge of labelling annotations, most conventional hyperspectral classifiers are tested and trained to use a solitary hyperspectral image cube, resulting in two challenges. Firstly, the existing classification models give reasonably close to perfect accuracy, but they are hard to extrapolate to other datasets. Second, hyperspectral datasets are not typically collected in the same scene; different data sets will contain distinct spectral bands. To address these concerns, a new framework for hyperspectral image stratification is proposed in this paper that requires training and analysing separately throughout heterogeneous hyperspectral datasets. Even though annotating hyperspectral data are beneficial, it has received very little attention from the hyperspectral community. The proposed zero-shot learning-based novel model maps the features between training and testing datasets by proposing a convolutional and word embedding model. The feature mapping module associates the training and testing set leading to classification. Experiments on benchmark hyperspectral datasets including Indian Pines (IP), Pavia University (PU), Kennedy Space Center (KSC) validate the performance of proposed cross-domain model.