Improved sub-pixel mapping method coupling spatial dependence with directivity and connectivity

Accurate land cover mapping by using coarse resolution imageries has been an attractive research topic. Sub-pixel mapping has been proven efficient for allocating sub-pixels within a mixed pixel. The most likely distribution can be determined on the condition of maximized spatial dependence. However, linear land cover like roads and rivers cannot be predicted efficiently because of weaker spatial dependence between and within mixed pixels. To obtain more accurate classification at the sub-pixel scale, an improved sub-pixel mapping method by combining spatial dependence with directivity and connectivity of linear land cover was proposed. Central line of linear land cover was extracted from fraction images to provide site-specific information. Discriminated allocation targets were accordingly designed: both connectivity and directivity were considered as important auxiliary information for allocating linear land cover, whereas only maximized spatial dependence is required for other classes. Then, simulated annealing arithmetic (SAA) was applied to optimize sub-pixel allocation. The method was evaluated visually and quantitatively with the accuracy indices. Compared with the model that considers only spatial dependence, SPM HIIPD method, attraction model and hard classifier (MLC), the improved method can increase classification accuracy at the sub-pixel scale with both simulated imageries and partial SPOT remotely sensed imagery.