Distributed algorithms for unmixing hyperspectral data using nonnegative matrix factorization with sparsity constraints

The ability to examine and extract the sources of data from hyperspectral images has become more and more important as the amount of data collected increases. Recent research has yielded better and better algorithms for unmixing this data to provide more accuracy. One such algorithm is Nonnegative Matrix Factorization which aims to approximate the sources of the known end result. An issue with current approaches is they are designed to be run sequentially and can be very computationally expensive. In this paper, ways of improving the performance of Sparse Nonnegative Matrix Factorization algorithms are introduced by utilizing distributed computing over a cluster of computers. The goal was to find ways of maximizing the throughput of a known algorithm without worrying about the accuracy of the algorithm itself (as this is shown through separate investigation). This was accomplished by testing out technologies such as MPI, POSIX threads and the OpenMP library. The aim was to compare and contrast different methods and find out what might be the optimal solution to allow for large data sets.