GPU implementation for hyperspectral image analysis using Recursive Hierarchical Segmentation

Graphics processing units (GPUs) have recently emerged as a promising parallel architecture considering the rapid growth of its computational power in comparison with uniprocessors. In this paper, we investigate a GPU parallel implementation of Recursive Hierarchical Segmentation technique (RHSEG). RHSEG is an object-based image analysis (OBIA) segmentation technique which involves region growing and spectral clustering (non-adjacent region growing). OBIA is becoming more popular compared to traditional pixel-based image analysis because of its efficiency with high spatial resolution images. The proposed parallel RHSEG algorithm was implemented using NVidia´s compute device unified architecture (CUDA). The experiments conducted show that the parallel GPU RHSEG achieved an average processing speedup of 3.5 times over the sequential CPU implementation.