Deep Feature Learning for Hyperspectral Image Classification and Land Cover Estimation

The differences in spatial sampling between field measurements and remote-sensing imagery can hinder the exploitation of contemporary data. When the field-based sampling is higher than airborne and spaceborne imagery, each pixel is naturally associated with multiple pixels due to the multiplexing of the reflectances of different materials. To address this scale inconsistency, we propose the introduction of the multi-label classification framework where classifiers are trained to predict multiple labels per pixel. Furthermore, instead of relying on raw hyperspectral measurements for the classification process, we investigate the Stacked Sparse Autoencoders framework, an example of a deep learning network, for descriptive feature extraction. To validate the merits of the proposed scheme, we consider real data from the Hyperion instrument on-board the EO-1 and NYC land cover data from 2010.