A Parallel Differential Box-Counting Algorithm Applied to Hyperspectral Image Classification

Abstract

In this letter, spatial information through fractal measures is adopted to combine with the spectral information to improve the land cover classification. The spectral features alone and later combined with texture features, using MODIS/ASTER airborne simulator imagery, were fed into a neural classifier. Classification performance was evaluated by a confusion matrix measured by overall accuracy and kappa coefficient. In particular, a parallel differential box-counting (DBC) (PDBC) algorithm for fractal estimation was implemented on a multicore PC. The computation efficiency was ensured through the use of PDBC algorithm which is much faster than that of the original DBC. Furthermore, multicore processors offer great potential for speeding up the computation by partitioning the load among the cores. Multithreading technique is adopted to fully explore its multicore capability. Experimental results demonstrate that the proposed approach provides substantial improvements in classification accuracy while requiring much less computation time without extra hardware resources.