Triplet-Watershed for Hyperspectral Image Classification

Hyperspectral images (HSI) consist of rich spatial and spectral information, which can potentially be used for several applications. However, noise, band correlations high dimensionality restrict the applicability such data. This is recently addressed using creative deep learning network architectures as ResNet, SSRN, A2S2K. last layer, i.e classification remains unchanged taken to softmax classifier. In this article, we propose use a watershed Watershed classifier extends operator from Mathematical Morphology classification. its vanilla form, does not have any trainable parameters. novel approach train networks obtain representations suitable The exploits connectivity patterns, characteristic HSI datasets, better inference. We show that exploiting characteristics allows Triplet-Watershed achieve state-of-art results in supervised semi-supervised contexts. These are validated on Indianpines (IP), University Pavia (UP), Kennedy Space Center (KSC) Houston (UH) relying simple convnet architecture quarter parameters compared previous state-of-the-art networks. source code reproducing experiments supplementary material (high resolution images) available at https://github.com/ac20/TripletWatershed Code.