Utilizing Parallel Networks to Produce Sub-Pixel Shifted Images With Multiscale Spatio-Spectral Information for Soft-Then-Hard Sub-Pixel Mapping

The distribution information of the land-cover classes in remote sensing image can be explored by sub-pixel mapping (SPM) technique. The soft-then-hard sub-pixel mapping (STHSPM) has become an important type of SPM method. The sub-pixel shifted images (SSI) from the same area can be utilized to improve the mapping result. However, the type of information in the fine SSI is insufficient, and the SSI-based STHSPM results are affected. To solve this problem, utilizing parallel networks to produce subpixel shifted images with multiscale spatio-spectral information (SSI-MSSI) for STHSPM is proposed. In SSI-MSSI, the fine SSI with multi-scale information and spatio-spectral information are obtained, respectively, from parallel networks, namely the multiscale network and spatio-spectral network. The multiscale network is spectral unmixing followed by mixed spatio attraction model and the spatio-spectral network is projected onto convex sets super-resolution followed by spectral unmixing. There two different kinds of fine SSI are integrated by appropriate weight parameter to produce the fine fractional images. Class allocation method then allocates the class labels into to each sub-pixel by the predicted value from the integrated fine fractional images. Three remote sensing images are tested to show that the proposed SSI-MSSI produces more accurate mapping results than the existing SSI-based STHSPM in the literature. In the quantitative accuracy assessment, the SSI-MSSI shows the best performance with the percentage correctly classified of 99.09% and 74.07% in the experimental results.