An unmixing framework to improve class accuracies using detected high importance local regions

Image Classification techniques are aimed at improving the class accuracies which are affected by the occurrence of mixed pixels in the remotely sensed data. Improving the labeling accuracies for the mixed pixels regions will increase the global class accuracies. Spectral unmixing has been used to decompose the mixed pixel regions into its constituent endmembers, and a corresponding fractional abundance for each endmember. The unmixing approaches are approximated based on spectral behavior, but ignore the spatial neighborhood. The data values at the pixel along with its spatial neighborhood are good indicators of the image characteristics including atmospheric conditions and need to be considered. In the current research, we propose a spatio-spectral framework that improves the classification accuracy and demonstrates its utility by improving the labels of the detected mixed regions for MODIS data and validated with AWIFS (APLULC 2005) derived land cover dataset.