AbstractSubpixel mapping (SPM) of a hyperspectral image (HSI) allocates land cover fractions or discrete abundances in original pixels, so that the resolution of the HSI label map becomes finer by dividing the mixed pixel to subpixels. Most of existing SPM approaches have the limitation of unlearnable spatial prior in HSIs and nonintegrated frameworks.
In this paper, we present an unsupervised Bayesian subpixel mapping network for hyperspectral images.
An end-to-end unified SPM network with an encoder-decoder architecture is designed to incorporate the fully convolutional neural network (FCNN) with the deep image prior and the forward models to effectively estimate the subpixel labels.
The proposed approach is tested on a benchmark real HSI dataset, in comparison with several other SPM methods. The results demonstrate that the proposed method is more effective for SPM of HSIs with higher numerical accuracies and more accurate visual maps of subpixel labels.