Sub-Pixel Mapping Based on Conditional Random Fields for Hyperspectral Remote Sensing Imagery

Sub-pixel mapping is a useful technique for providing land-cover information at the sub-pixel scale by the use of the input fraction image at a coarse resolution. Some sub-pixel mapping algorithms with strict consideration of the abundance constraint have difficulty in obtaining a satisfactory performance in sub-pixel mapping since the fraction image obtained by spectral unmixing always contains errors. In this paper, in order to make full use of the input fraction image and alleviate the effect of fraction errors, a sub-pixel mapping algorithm based on conditional random fields (CRFSM) is proposed for hyperspectral remote sensing imagery. The CRFSM algorithm fuses the local spatial prior at the fine scale and the downscaled coarse fraction at the coarse scale by potential functions to obtain more detailed land-cover distribution information. The local spatial prior models the local spatial structure to obtain the local land-cover features at the fine scale. The downscaled coarse fraction considers the fraction values to maintain the holistic land-cover features at the coarse scale. In addition, the abundance constraint is considered as a soft constraint by the probability class determination strategy in the CRFSM algorithm, to help with the class label determination of sub-pixels and alleviate the effect of the fraction errors and noise. The experimental results with two synthetic hyperspectral images and a real Nuance hyperspectral image show that the proposed sub-pixel mapping algorithm has a competitive performance in both the quantitative and qualitative evaluations, compared with the other state-of-the-art sub-pixel mapping algorithms.