Parallel implementation of multiple kernel self-organizing maps for spectral unmixing

Spectral unmixing algorithms are commonly used in processing of hyperspectral images to identify the elemental components, called end-members, and their corresponding information in each pixel of the image. However, these algorithms are computationally intensive and can become a bottleneck for remote sensing hyperspectral image processing, especially in large aerial imagery processing centers. This paper, explores the use of massive parallel processing graphical processing unit to speed up the multi kernel self-organizing map (MKSOM) unmixing algorithm. MKSOM is based on artificial neural networks, which makes it suitable to be efficiently parallelized. Two real benchmark hyperspectral images; AVIRIS Cuprite and Brullus are used to evaluate the performance of the parallel algorithm. The experimental results show that the proposed implementation is appropriated for real-time hyperspectral remote sensing applications due to a very small worst case parallel execution time (0.83 s when the number of classes is less than 9) which makes it feasible to be integrated as on-board processing on any Hyperspectral remote sensors. Our parallel technique achieved a significant speedup compared with a multi-threaded CPU implementation applied on the same hyperspectral image. The results showed a speedup of 93.46 x for SOM size of 256 and trained for 100 epochs on medium-sized HSI such as AVIRIS Cuprite.