A parallel differential box counting algorithm applied to hyperspectral image classifications

Hyperspectral images with hundreds of narrow spectral channels are currently available and instruments with thousands of spectral bands are under development. It is necessary to develop techniques and models for efficiently processing large volume of remote sensing images. Multi-core processors present an opportunity for speeding up the computation by partitioning the load among the cores. As multi-core processor systems become more and more widespread, the demand of efficient parallel algorithms also propagates into the field of remote sensing images processing. Classification of land cover types in a hyperspectral image is demonstrated in this study. A dynamic learning neural network (DLNN) is utilized as a supervised classifier. To get better classification accuracy, texture information is extracted and combined with the spectral information. Fractal dimension, the texture information applied, is estimate by a differential box-counting (DBC) technique. The original DBC is inefficient because the fractal dimension is evaluated sequentially. In this study, a parallel DBC is proposed and implemented on a multi-core PC to improve its efficiency. To fully explore its multi-core capability, multi-threading technique is adopted. Experimental results reveal that the improvement in computation time of the parallel DBC is depended on the ratio of window size M and grid size s. In addition, further improvement provided by multi-threading techniques is linearly proportion to the number of cores.